Years and Dollars

Introduction

Personal finance keeps one ledger and medicine keeps another. The first counts dollars; the second counts years. This paper argues that they are a single ledger, and that its most consequential entries are written in the decades between leaving school and reaching retirement.

The link runs in both directions. Wealth buys time. Across the U.S. income distribution, higher income is associated with longer life at every percentile ; in a national cohort followed from the mid-1990s, the mortality gap by wealth was larger than the gaps by education, occupation, or income ; low wealth in later middle age is associated with both death and disability ; and a sudden collapse in wealth is followed by a sharp rise in mortality . Time, in turn, builds wealth. Savings compound only over the years a saver lives, retirement income arrives only while a retiree survives, and a death in the household removes an earner, imposes costs, and often forces the sale of assets.

Both directions bear on the racial wealth gap, which Derenoncourt and colleagues describe as the largest economic disparity between Black and White Americans and one of the most persistent . In 2022 the typical White family held $285,000 in net worth and the typical Black family $44,900 . Measured per person, White Americans held about six times as much wealth as Black Americans in 2020; the gap stopped narrowing around 1950 and has widened since the 1980s, as gains on assets went mostly to White households . Black Americans also live shorter lives: in 2024, life expectancy at birth was {{v:e0_b}} years for the non-Hispanic Black population and {{v:e0_w}} for the non-Hispanic White population . It is natural to ask how much of the first gap the second explains. Our answer is that the part we can measure directly in net worth is small: about {{v:comp_share}} of the log wealth gap through the age structure of the adult population and about {{v:work_share}} through shorter working lives. Through channels that net worth does not record, the amounts are larger: a projection model built on the Social Security Trustees’ mortality assumptions puts the cost of shorter lives at {{v:pj_pen_m}} of Social Security wealth for a Black man entering work today, and at {{v:pj_lost_pen}} of expected lost earnings for a Black couple aged 30.

This paper makes three contributions and adds a review. First, after synthesizing the evidence on the two-way relationship between wealth and longevity (Section 2), we use the 2024 life tables to locate the Black–White survival gap by age and to quantify four accounting channels through which it reaches wealth: the age composition of the population, the annuity structure of retirement income, the loss of partners within households, and the length of working lives (Section 3). Second, we build an open projection model that puts the channels net worth does not record in dollars: Social Security wealth with spousal, survivor, disability, and children’s benefits, the economics of losing a partner, and the timing and size of bequests, with forecasts to 2075 under the Social Security Trustees’ mortality assumptions (Section 4). Third, we build a microsimulation of household wealth, calibrated to the 2022 Survey of Consumer Finances, tested by running it forward from the 1989 and 1992 surveys, and run for {{v:ws_hh_per}} households per path, that asks what would close the gap, how quickly, how much each lever contributes, and what closing the gap would add in dollars and in years of life (Section 5). Its headline specification, which fits the survey and its history better than the alternative, lets each group’s saving rate differ and finds Black households saving less at equal income; we therefore treat equal saving as a lever of its own, separate from equal earnings, returns, and survival, and report one saving schedule for both groups as a sensitivity. Like structural models, the microsimulation finds that earnings parity is the dominant lever and that one-time transfers fade , and an independent re-implementation with separate code reproduces its inputs and its main conclusions (Section 5.8). Finally, we review the evidence on the quality of the later years (Section 6), on stress spending, one everyday route by which strain can become debt and debt can become strain (Section 7), and on other wealth gaps that similar mechanisms may widen: those of women, Hispanic Americans, American Indian and Alaska Native people, people with disabilities or mental health conditions, and households harmed by gambling, penalty fees, or exclusion on grounds of their values (Section 8). Section 9 sets out threats to validity and Section 10 concludes.

A note on terms. “Black” and “White” refer to the non-Hispanic Black and non-Hispanic White populations as classified by the National Center for Health Statistics and the Survey of Consumer Finances. Racial differences in survival and wealth reflect differences in conditions, opportunities, and exposures, not biology .

Longevity and wealth: a two-way street

Wealth buys time

The income gradient in U.S. mortality is steep and continuous. Linking 1.4 billion person-year tax records to death records, Chetty and colleagues found that life expectancy rises with income throughout the distribution; in 2001–2014 the gap in life expectancy at age 40 between the richest and poorest 1% was 14.6 years for men and 10.1 years for women, and it widened over the period . The National Academies projected that men born in 1960 who reach 50 can expect to live to about 89 if they are in the top fifth of earners and to about 76 if they are in the bottom fifth, a gap of more than 12 years, up from about 5 years for men born in 1930 .

For mortality in middle and later life, wealth is an even sharper predictor than income, perhaps because it measures accumulated security rather than a single year’s flow. In the Midlife in the United States cohort, the mortality disparity by wealth was larger than the disparities by education, occupation, income, or childhood socioeconomic status: after age 65, adults with no assets had an estimated 40% chance of surviving to 85, against 71% for those with at least $300,000 (in 1995 dollars), a difference that narrowed to 51% versus 70% after full adjustment for confounders . In the Health and Retirement Study, Americans aged 54–64 in the lowest wealth quintile had a 17% risk of death and a 48% risk of disability over ten years, against 5% and 15% in the highest quintile . The gradient is unusually steep in the United States: among 73,838 adults aged 50–85 in American and European cohorts, survival in the wealthiest American quartile appeared similar to survival in the poorest quartile of northern and western Europe .

Changes in wealth matter as well as levels. Among Americans aged 51–61 at entry to the Health and Retirement Study, a negative wealth shock, defined as losing 75% or more of net worth over two years, was associated with an adjusted hazard ratio for death of 1.50, and asset poverty at baseline with 1.67; death rates were 64.9 per 1,000 person-years after a shock against 30.6 for those whose wealth stayed positive .

Several mechanisms carry wealth into survival: access to care, healthier housing and neighborhoods, the ability to absorb a shock without forgoing treatment, food, or rent, and relief from the physiological cost of chronic insecurity. McEwen’s account of allostatic load describes how the stress mediators that protect the body in the short run damage it when they are activated chronically . Financial insecurity can be such a chronic stressor; Section 7 reviews the evidence linking debt with stress and health.

Time builds wealth

The reverse direction is stated less often, but it is just as mechanical. Wealth is a stock accumulated from flows over time: earnings saved, returns compounded, debts repaid. Each additional working year adds a year of earnings and contributions, and each additional year of life lengthens the horizon over which assets compound. Social Security retirement benefits are an annuity: they are paid only while the beneficiary lives, and in December 2024, 55.0 million Americans aged 65 or older received Social Security benefits . Because lifespans have grown faster for high earners, the National Academies projected that men born in 1960 in the top fifth of earners will receive $132,000 more in lifetime benefits from the major entitlement programs (Social Security, Medicare, Medicaid, and Supplemental Security Income), in present value at age 50, than men in the bottom fifth .

Health shocks that precede death are also wealth shocks. Three years after a hospital admission, insured adults admitted at ages 50–59 were 11 percentage points (15%) less likely to be employed and earned $9,000 a year less (20% of their prior earnings), a loss far larger than the $1,400 annual rise in their out-of-pocket medical spending . Deaths impose further costs on survivors: lost earnings, funeral expenses, caregiving time, and, often, the sale of a home or business to settle an estate.

History shows how large the time channel can be. Karger and Wray estimate that the average White man born in 1900 earned 2.6 times as much over his lifetime as the average Black man, nearly twice the cross-sectional earnings gap, because 48% of Black men born in 1900 died before age 30, against 26% of White men. Convergence in mortality halved the Black–White welfare gap between the 1900 and 1920 birth cohorts; for cohorts born between 1920 and 1970, the gap stagnated as differences in income and life expectancy remained large .

The racial wealth gap and the lifespan gap

Two gaps

Wealth. The 2022 Survey of Consumer Finances put median net worth at $285,000 for White families and $44,900 for Black families; mean net worth was $1,367,200 and $211,500 . The typical Black family thus held {{v:obs_ratio_med}} of the wealth of the typical White family. The gap’s origins are historical and structural: vastly different starting conditions after Emancipation ; exclusion from mortgage credit and homeownership, including through the 1930s Home Owners’ Loan Corporation maps, which reduced homeownership, house values, and rents and increased racial segregation in the affected neighborhoods for decades ; and continuing disparities in income, returns, and access to credit. Differences in individual choices such as education, saving, or financial literacy do not close it . Family transfers sustain it. In the 2019 survey, 29.9% of White families had received an inheritance or gift, against 10.1% of Black families; 17.1% versus 6.0% expected one; 71.9% versus 40.9% could get $3,000 from family or friends in an emergency; and among middle-aged families, 65% of White and 44% of Black families held a retirement account .

Lifespan. In 2024, life expectancy at birth was {{v:e0_b}} years for the non-Hispanic Black population and {{v:e0_w}} for the non-Hispanic White population, a gap of {{v:gap}} years. The gap was {{v:gap_m}} years for men ({{v:e0_bm}} versus {{v:e0_wm}}) and {{v:gap_f}} for women ({{v:e0_bf}} versus {{v:e0_wf}}) . Table {{n:tab-survival}} summarizes the quantities used below; all are computed from the complete 2024 period life tables.

Where the survival gap arises

Figure {{n:fig-survival}} plots survivors by age. The curves separate early and widen through middle age. By 65, {{v:p65_w}} of a White birth cohort is still alive against {{v:p65_b}} of a Black cohort, and half of each cohort has died by about age {{v:half_w}} and {{v:half_b}} respectively. After 65 the remaining difference is comparatively small: a Black 65-year-old can expect {{v:e65_b}} more years and a White 65-year-old {{v:e65_w}}, a difference of {{v:e65_gap}} years.

To locate the gap precisely, we decompose the difference in life expectancy at birth by age at death using the method of Arriaga . For single years of age \(x\), with survivors \(l_x\), person-years lived \(L_x\), and person-years remaining \(T_x\), and superscripts \(B\) and \(W\) for the two populations, the contribution of deaths at age \(x\) to \(e_0^W - e_0^B\) is

with the open age interval at 100 handled by a single term. The contributions sum exactly to the gap. Of the {{v:gap2}}-year difference, {{v:arr_before65}} years ({{v:arr_before65_pct}}) arise from deaths before age 65 and {{v:arr_2264}} years ({{v:arr_2264_pct}}) from deaths at working ages, 22–64 (Figure {{n:fig-arriaga}}). Deaths at 45–64 alone contribute {{v:arr_4564}} years ({{v:arr_4564_pct}}); infant deaths contribute {{v:arr_inf}} years. Above 85 the contribution is slightly negative ({{v:arr_85}} years): recorded death rates at the oldest ages are slightly lower for Black than for White Americans, which may partly reflect selective survival and the quality of data at those ages, where the life tables rely on Medicare records and smoothed death rates . The decomposition also depends on which population is taken as the starting point: reversing the roles of the two populations in Equation 1 attributes {{v:arr_before65_rev}} years ({{v:arr_before65_rev_pct}}) to deaths before 65 and {{v:arr_85_rev}} years to deaths above 85, and the average of the two directions attributes {{v:arr_before65_avg_pct}} to deaths before 65.

This age pattern is the signature of what Geronimus called weathering: the cumulative physiological toll of chronic social and economic adversity, which ages bodies early . In national data, Black adults had higher allostatic load scores than White adults at all ages, most of all at 35–64, and the difference was not explained by poverty . Wealth is part of the story: in the Health and Retirement Study, the far lower asset holdings of Black adults affected not only their financial well-being but their survival, independent of income and education . The composition of wealth matters too: among Black and White adults aged 25 and older, savings, stock ownership, and homeownership were associated with better health and debt with worse health, and the health returns to assets differed by race, although the authors found no evidence that debt was more harmful to Black than to White adults . Bereavement compounds the burden: in two national surveys, Black Americans were more likely than White Americans to have lost a mother, a father, or a sibling by midlife, and a child or a spouse during adulthood .

How shorter lives reach wealth

Related work. Decompositions of the Black–White gap in life expectancy by age and cause of death are standard; Harper and colleagues used one to show that the gap narrowed after 1993 because of relative declines in Black mortality from homicide, HIV, unintentional injuries, and, among women, heart disease . A separate literature shows how mortality differences change what Social Security is worth: when lower earners die younger, much of the progressivity of the benefit formula is undone , and counting Social Security wealth alongside the assets that surveys measure reduces measured wealth inequality . The race-specific evidence on Social Security that we cite is older . Our contribution is to update the accounting to the 2024 life tables, to put each survival channel on the scale of observed net worth where the data allow, and to say plainly which channels cannot be put on that scale.

We now ask how the survival differences in Table {{n:tab-survival}} enter wealth when everything else is held equal. Each calculation gives both populations identical earnings, saving, returns, and benefit rules and lets only survival differ. These are accounting exercises, not causal estimates. Their purpose is to bound how much of the wealth gap the survival gap can carry through each channel.

Channel 1: age composition. Wealth rises with age into the late sixties (Figure {{n:fig-composition}}A). A population whose members die younger has fewer adults at the ages when wealth peaks. To isolate this effect, we weight the 2022 age-wealth profile of all U.S. families, \(w(a)\) , by each population’s adult person-years:

Because Black survival is lower, Black adult person-years tilt toward younger ages (Figure {{n:fig-composition}}B), and the age-standardized wealth index is {{v:comp_med_pct}} lower using median wealth by age and {{v:comp_mean_pct}} lower using mean wealth by age. Against the observed Black-to-White ratios of median wealth ({{v:obs_ratio_med}}) and mean wealth ({{v:obs_ratio_mean}}), this is {{v:comp_share_med}} of the log gap with medians and {{v:comp_share_mean}} with means; across the metrics in Table {{n:tab-compsens}}, which also start adulthood at 18 and measure the gap in levels, the share lies in the range {{v:comp_range}}. Because the weights come from period life tables, the calculation isolates the effect of mortality on age structure; the actual Black and White adult populations also differ in age structure because of fertility and migration. The national age-wealth profile is itself shaped by differential survival, since wealthier people live longer. The compositional channel is real but small.

Channel 2: the annuity structure of retirement income. Social Security retirement benefits are paid for as long as the beneficiary lives. For a person alive at 22, expected years of life at 67 and older are {{v:ret_b}} in the Black population and {{v:ret_w}} in the White population, {{v:ret_pct}} fewer; for men the figures are {{v:ret_bm}} and {{v:ret_wm}}, {{v:ret_m_pct}} fewer. To translate years into money, consider two workers with identical earnings who pay payroll tax from 22 through 66 while alive and receive the same benefit from 67 while alive. The ratio of the expected present value of benefits to that of contributions is

and the Black-to-White ratio of \(\mathrm{MW}\) isolates survival. At a 2% real discount rate, a Black worker’s expected retirement benefits per dollar of contributions are {{v:ss_all_1}} lower than those of a White worker with the same earnings, and {{v:ss_men_1}} lower for men (Table {{n:tab-ss}}). The result barely moves with the discount rate.

Retirement benefits are not the whole of Social Security. The program also pays disability and survivor benefits, and its benefit formula replaces a larger share of earnings for lower earners. In a 2003 microsimulation of the full program for people born from 1931 to 1964, the General Accounting Office (now the Government Accountability Office) found that Black and Hispanic Americans as groups tend to receive more in benefits relative to taxes than White Americans because they have lower lifetime earnings and higher disability rates, and that these aggregate differences are small compared with the differences between high and low earners within each group . That comparison does not hold earnings equal, so it should not be read as offsetting the {{v:ss_all_1}} shortfall at equal earnings in Table {{n:tab-ss}}. Examined through the retirement component alone, which depends heavily on life expectancy, Black Americans fare worse than White Americans, as the GAO also noted . Section 4 models the disability and family benefits explicitly.

Channel 3: the loss of a partner. Households pool earnings, and a death before retirement removes an earner. If two partners aged 30 face independent survival risks, the probability that at least one dies before 67 is \(1 - S_m(30,67)\,S_f(30,67)\), where \(S\) denotes survival between the two ages in the sex-specific life tables. For Black couples that probability is {{v:cpl30_b1}}; for White couples, {{v:cpl30_w1}}. For couples aged 40, it is {{v:cpl40_b1}} and {{v:cpl40_w1}}. These figures assume partners of the same age whose survival is independent and follows the population life tables, which include unmarried people, who tend to die younger than married people; partners’ risks are also correlated, so the levels are approximate. Each such death is a compound wealth shock: lost earnings, funeral and end-of-life costs, and, for a surviving parent of young children, years of single-income saving.

Channel 4: working lives. For a person alive at 22, expected years lived between 22 and 66 are {{v:work_b}} in the Black population and {{v:work_w}} in the White population, {{v:work_pct}} fewer ({{v:work_m_pct}} for men). Holding annual earnings equal, expected lifetime labor income falls by the same proportion; if net worth scaled one-for-one with years worked, this channel would account for about {{v:work_share}} of the log median wealth gap. This channel was enormous for cohorts born in 1900 . Today it is modest, but it is concentrated in the deaths at 45–64 that Figure {{n:fig-arriaga}} highlights, when earnings and savings are near their peak.

Channels we do not quantify. Health shocks before death reduce employment and earnings . Medical debt is more common among Black households: 28% reported it in a 2018 Census Bureau survey summarized by the Consumer Financial Protection Bureau, against 17% of White households . And early, repeated bereavement strains the same family networks that provide emergency help and transfers .

Summary of the channels

Table {{n:tab-channels}} collects the estimates. They support three conclusions.

First, shorter lives explain little of the measured wealth gap directly. The two channels that can be put on the scale of observed net worth are small: the compositional effect is about {{v:comp_share}} of the log gap ({{v:comp_range}} across metrics), and shorter working lives account for about {{v:work_share}} if wealth scaled one-for-one with years worked. The gap is primarily a product of history and of continuing differences in income, inheritance, homeownership, and access to credit .

Second, survival differences bear on forms of wealth and risk that net worth does not record. At equal earnings, retirement benefits per contribution dollar are {{v:ss_all_1}} lower for Black workers, and a Black couple aged 30 has a {{v:cpl30_b}} rather than {{v:cpl30_w}} chance of losing a partner before 67, under the assumptions above. Social Security wealth is excluded from the survey measure of net worth , and the partner calculation is a probability, so neither can be read as a share of the observed gap. Section 4 builds a projection model that puts these channels, and bequests, in dollars.

Third, the arrows run both ways. Lower wealth predicts earlier death , and earlier death lowers what a household can accumulate. Chronic stress is a plausible common contributor to both gaps: it is associated with weathering and with spending and borrowing under strain (Section 7). The evidence reviewed here does not show that stress drives the racial wealth gap, whose main causes lie in the history and institutions cited above.

Projecting the channels net worth does not record

Section 3 put two survival channels on the scale of measured net worth and left three as ratios or probabilities: Social Security wealth, the loss of a partner, and bequests. This section builds an open projection model for those three. It asks what the Black–White survival gap is worth in dollars through each channel, for whom, and how the amounts change as death rates fall. Like Section 3, the model is accounting, not causal inference: it changes survival, and where stated earnings, while holding behavior fixed.

A projection model

Mortality. We start from the 2024 life tables for non-Hispanic Black and White men and women and carry them forward with the improvement in death rates by age and sex that the 2026 Social Security Trustees Report projects to 2100 :

where \(q^{g}_{x,t}\) is the probability that a member of group \(g\) aged \(x\) in year \(t\) dies before reaching \(x+1\). Under the Trustees’ intermediate assumptions this holds the Black–White gap in log death rates at its 2024 value at every age, so the gap measured in years narrows as death rates fall; their low-cost and high-cost assumptions bound the projection. A second scenario lets the gap close linearly to zero by 2050. Each cohort lives through the projected rates year by year, and the model reproduces the 2024 life expectancies exactly.

Earnings and benefits. Each worker earns the age profile of the Social Security Administration’s medium scaled worker, whose average over the highest 35 years equals the national average wage index ({{v:pj_awi}} in 2024) , scaled by the group’s mean earnings relative to all earners aged 25 to 64 in the Current Population Survey : {{v:pj_lv_wm}} for White men, {{v:pj_lv_bm}} for Black men, {{v:pj_lv_wf}} for White women, and {{v:pj_lv_bf}} for Black women. A typical-worker variant uses median earnings instead. Benefits follow the law: the average of the highest 35 years of earnings up to the taxable maximum (fewer years for deaths before 62), the 2026 benefit formula, which is indexed to 2024 wages, claiming at the full retirement age of 67, a spousal benefit of up to half the other spouse’s benefit, and a survivor benefit equal to the deceased worker’s benefit . We report two measures. The retirement measure, that of version 2.0, sets retirement, spousal, and survivor benefits from 67 against the 10.6% retirement and survivors share of the payroll tax paid while alive. The full measure adds the benefits that pay out on disability and on early deaths: disabled-worker benefits; for a disabled worker’s family, 50% of the worker’s benefit for each child and for a spouse caring for a child under 16; for a deceased worker’s family, 75% for each child under 18 (19 in high school) and for a surviving parent under 60 caring for a child under 16; and the reduced widow(er)’s benefit from 60, subject to the earnings test; with the family maximum applied throughout . It sets them against the full 12.4% tax in years not spent on disability. Disability rates are the Social Security Administration’s December 2024 counts of disabled workers by sex and age divided by the population of that sex and age, times each group’s receipt rate relative to everyone of its sex and age band in the 2025 Current Population Survey microdata, with standard errors from its replicate weights . At 60 to 66, {{v:pj_di_b60_m}} of Black and {{v:pj_di_w60_m}} of White men receive them, and {{v:pj_di_b60_f}} of Black and {{v:pj_di_w60_f}} of White women. A disabled worker receives the average benefit of 2024 awards for that sex, scaled by the group’s ratio of reported benefits in the survey ({{v:pj_di_pia_bm}} a month for Black and {{v:pj_di_pia_wm}} for White men), and the ages of children in each family come from the same microdata. Claiming at 62 is a sensitivity. Amounts are in 2024 dollars, benefits are those scheduled under current law, and present values use a 2% real discount rate.

Wealth and costs. Net worth by race, age, and marital status comes from the 2022 Survey of Consumer Finances microdata, with standard errors from its 999 replicate weights and five implicates . A death costs the median funeral, {{v:pj_funeral}} , and, from age 65, the out-of-pocket medical spending of the last five years of life measured in the Health and Retirement Study, a mean of {{v:pj_oop_mean}} and a median of {{v:pj_oop_med}} , converted to 2024 dollars with the consumer price index . Couples are a woman and a man two years older whose lives are independent. The code and every input are in the replication files.

Social Security wealth

At equal earnings, survival alone means that a Black man entering work in 2024 can expect {{v:pj_mw_eq_gap_m}} less in retirement benefits per dollar of contributions than a White man entering at the same time (Table {{n:tab-ssw}}). With 2024 death rates held fixed the model gives {{v:pj_mw_eq_gap_m24}}, the figure in Table {{n:tab-ss}}; projected improvement narrows the gap in years and so the gap in money’s worth. At the two groups’ actual earnings the difference nearly disappears ({{v:pj_mw_bm}} against {{v:pj_mw_wm}}), because the benefit formula replaces a larger share of lower earnings. Counting disability and family benefits against the full tax closes the gap at equal earnings almost entirely ({{v:pj_mwf_eq_bm}} in benefits per dollar of tax for a Black man against {{v:pj_mwf_eq_wm}} for a White man), and at actual earnings a Black man now expects more per dollar than a White man ({{v:pj_mwf_bm}} against {{v:pj_mwf_wm}}), because Black workers are more likely to draw disability benefits and to leave children who draw survivor benefits. The progressive formula and these benefits mask the survival penalty rather than removing it: lower lifetime earnings buy a higher replacement rate, higher disability and earlier deaths draw more of the program’s insurance, and shorter lives take part of it back. Giving Black workers White survival while keeping their own earnings, disability rates, and families isolates the penalty. Counting all benefits, for a Black man entering work in 2024 it is {{v:pj_pen_m}} of expected Social Security wealth in present value ({{v:pj_pen_m0}} undiscounted; {{v:pj_pen_m_ret}} counting retirement benefits only, the measure of version 2.0); for a Black woman, {{v:pj_pen_f}} ({{v:pj_pen_f0}}); and for a Black couple, {{v:pj_pen_cpl}} ({{v:pj_pen_cpl0}}). Had everyone claimed at 62, the man’s penalty would be {{v:pj_pen_m62}}. The chance of living long enough to recover one’s contributions in benefits tells the same story: at equal earnings it is {{v:pj_rec_eq_bm}} for Black men and {{v:pj_rec_eq_wm}} for White men, and at actual earnings {{v:pj_rec_bm}} and {{v:pj_rec_wm}}.

Counting Social Security wealth

For the typical older Black household, Social Security wealth is worth about {{v:pj_ssw_x_b}} times its net worth; for the typical older White household, {{v:pj_ssw_x_w_text}}. Among households aged 65 to 74, median net worth was {{v:pj_nw_med_b}} for Black households (standard error {{v:pj_nw_med_b_se}}) and {{v:pj_nw_med_w}} for White households ({{v:pj_nw_med_w_se}}), in 2024 dollars, while the Social Security wealth of a typical-earning household of each group at 70 is {{v:pj_ssw_typ_b}} and {{v:pj_ssw_typ_w}} (Table {{n:tab-aug}}). Counting it raises the Black-to-White ratio of typical wealth from {{v:pj_ratio_nw_med}} to {{v:pj_ratio_aug_typ}} and of mean wealth from {{v:pj_ratio_nw_mean}} to {{v:pj_ratio_aug_mean}}, consistent with evidence that adding Social Security wealth reduces measured wealth inequality ; had every household claimed at 62, the typical ratio would rise to {{v:pj_ratio_aug_typ62}}. Because older Black households depend more on this asset, its survival penalty bears more heavily on them: at 70, shorter remaining lives cost Black households {{v:pj_pen70_typ}} to {{v:pj_pen70_mean}} of expected Social Security wealth, about {{v:pj_pen70_pct}} of it. This counts only the survival gap after 70; for workers who are 21 today the penalty also includes the higher chance of dying before benefits begin (Table {{n:tab-ssw}}). The Black estimates rest on {{v:pj_fam_b}} sampled families, hence their wide standard errors.

The loss of a partner

For couples aged 30 and 32 in 2024, the model projects that {{v:pj_ploss_b}} of Black couples and {{v:pj_ploss_w}} of White couples will lose a partner before the woman turns 67, and that {{v:pj_widf_b}} of Black women and {{v:pj_widf_w}} of White women will be widowed by then, spending on average {{v:pj_ywid67_b}} and {{v:pj_ywid67_w}} years widowed before 67 (Table {{n:tab-partner}}). Section 3’s {{v:cpl30_b}} and {{v:cpl30_w}} are higher because they hold 2024 death rates fixed; Table {{n:tab-partner}} also shows partners of the same age. The expected present value of the earnings a household loses when a partner dies before 65 is {{v:pj_lost_b}} for Black couples and {{v:pj_lost_w}} for White couples, whose higher earnings make each death more costly. With White survival, Black couples would lose {{v:pj_lost_bw}}: shorter lives add {{v:pj_lost_pen}} to their expected loss.

Survivor insurance replaces part of the loss, through three kinds of benefit. Children under 18 and a surviving parent caring for a child under 16 draw benefits while the children are young: their expected value is {{v:pj_kids_b}} for Black couples and {{v:pj_kids_w}} for White couples. A surviving partner may draw a reduced survivor benefit from 60, subject to the earnings test: {{v:pj_surv60et_b}} and {{v:pj_surv60et_w}}. And from 67 the survivor receives the larger of the two partners’ benefits, which is worth most when one partner earned much more than the other; Black partners’ benefits are lower and closer together ({{v:pj_pia_bf}} and {{v:pj_pia_bm}} a month) than White partners’ ({{v:pj_pia_wf}} and {{v:pj_pia_wm}}), so this part is worth {{v:pj_surv_b}} to Black couples and {{v:pj_surv_w}} to White couples. In all, survivor benefits are worth {{v:pj_surv_all_b}} to Black and {{v:pj_surv_all_w}} to White couples and replace {{v:pj_surv_share_b}} and {{v:pj_surv_share_w}} of their expected lost earnings. The benefits that pay out on deaths before 67 are larger for Black couples ({{v:pj_early_b}} against {{v:pj_early_w}}), and the extra survivor benefits that higher mortality brings Black couples, {{v:pj_surv_all_gain}}, offset {{v:pj_surv_all_offset}} of the {{v:pj_lost_pen}} that their shorter lives add to lost earnings. Social Security’s survivor insurance therefore pays less, and replaces less of the loss, where widowhood comes earliest. Disability benefits for the partners and their children add {{v:pj_di_cpl_b}} for Black and {{v:pj_di_cpl_w}} for White couples. Children’s ages for these families come from {{v:pj_fam_cpl_b}} Black and {{v:pj_fam_cpl_w}} White married couples in the survey microdata.

Bequests

Shorter lives change when children inherit more than how much. A child born in 2024 to Black parents aged {{v:pj_par_b}} expects to inherit from the last surviving parent at about {{v:pj_inh_age_b}}, and at about {{v:pj_inh_age_bw}} if the parents had White survival (Table {{n:tab-bequest}}). The expected estate, measured by the net worth of unmarried households of the parent’s age less funeral and end-of-life costs, is {{v:pj_est_typ_b}} at the median and {{v:pj_est_mean_b}} at the mean, and nearly the same with White survival ({{v:pj_est_typ_bw}} and {{v:pj_est_mean_bw}}). For a child of White parents aged {{v:pj_par_w}} it is {{v:pj_est_typ_w}} and {{v:pj_est_mean_w}}. The bequest gap is a gap in parents’ wealth, not in their lifespans. What shorter lives do change is the risk of losing both parents early: {{v:pj_orph45_b}} of Black children are projected to lose both parents before 45, against {{v:pj_orph45_bw}} with White survival and {{v:pj_orph45_w}} of White children, whose parents are older when they are born . These estimates rest on small samples of unmarried Black households ({{v:pj_fam_unm_b}} families aged 65 and older), and the wealth of living households overstates that of decedents, who are poorer on average .

Forecasts to 2075

Under the intermediate assumptions with the 2024 gap held in proportion, life expectancy rises for both groups and the gap in years narrows slowly: for men from {{v:pj_gm_2024}} years in 2024 to {{v:pj_gm_2050}} in 2050 and {{v:pj_gm_2075}} in 2075, and for women from {{v:pj_gf_2024}} to {{v:pj_gf_2050}} and {{v:pj_gf_2075}} (Figure {{n:fig-forecast}}). The penalties shrink more slowly still (Table {{n:tab-forecast}}). For Black men entering work in 2054, the Social Security penalty is {{v:pj_pen_m_2054}}, against {{v:pj_pen_m}} for those entering in 2024; across the Trustees’ low-cost and high-cost assumptions the 2024 figure ranges from {{v:pj_pen_m_lo}} to {{v:pj_pen_m_hi}}. The share of couples aged 30 who lose a partner before 67 falls from {{v:pj_ploss_b}} to {{v:pj_ploss_b_2054}} for Black couples who are 30 in 2054, and from {{v:pj_ploss_w}} to {{v:pj_ploss_w_2054}} for White couples. Falling mortality alone does not close these gaps. If Black death rates instead converged to White rates by 2050, the Social Security penalty for today’s 21-year-olds would nearly vanish ({{v:pj_pen_conv_m}}), because most of it arises at ages they would reach after 2050; couples aged 30 today would still bear {{v:pj_lost_conv}} of their {{v:pj_lost_pen}} partner-loss penalty, because the excess deaths before 2050 fall in their working years.

What the model adds. Through the channels that net worth does not record, the survival gap is worth tens of thousands of dollars to each Black household: {{v:pj_pen_f}} to {{v:pj_pen_cpl}} of Social Security wealth in present value, and {{v:pj_lost_pen}} of expected lost earnings for a couple, of which survivor insurance replaces about {{v:pj_surv_all_offset}}. These sums are large next to the median net worth of Black families, about {{v:pj_nw_b_all}} in 2024 dollars, and small next to the White–Black gap in it, about {{v:pj_gap_all}} . They confirm the conclusion of Section 3 from a second direction. The racial wealth gap is primarily a product of history and of continuing differences in income, inheritance, homeownership, and access to credit; shorter lives impose an additional, measurable tax on the assets that Black households depend on most, and the tax persists for as long as the survival gap does.

What it would take to close the gap

Sections 3 and 4 measured what shorter lives cost. This section asks the question in reverse: what would close the Black–White wealth gap, how quickly, and what closing it would add to the economy and to Black lives. We answer it with a microsimulation that starts from the households of the 2022 Survey of Consumer Finances and follows them and their descendants to 2100, simulating {{v:ws_hh_per}} households for each path, and checks it by running it forward from the surveys of 1989 and 1992. Like the rest of Part Ithis paper, the model is accounting under stated assumptions. It shows what each lever would do if it were achieved, not whether any policy would achieve it. Its headline specification lets each group’s base saving rate differ; a sensitivity with one saving schedule for both groups is the only specification under which every single lever and combination was run, and results that come from it are labeled so.

A microsimulation of household wealth

Households. Each simulated household begins as a non-Hispanic Black or White household drawn, with its survey weight, from the 2022 Survey of Consumer Finances . Each year its net worth changes as

where \(Y_{i,t}\) is income other than interest, dividends, and capital gains, set by the household’s rank \(p_{i}\) in the national income distribution, an age profile, and real wage growth; \(s_{g}(p_{i})\) is the share of that income saved, which rises with income rank from a base rate fitted for each group \(g\); \(\phi_{g}\) is the share of income given as support to relatives and friends outside the household, measured in the survey at {{v:ws_sup_b}} of income for Black and {{v:ws_sup_w}} for White households under 67 ({{v:ws_supgive_b}} and {{v:ws_supgive_w}} give any) ; \(c_{a}\) is consumption out of positive wealth, which after 67 draws wealth down over the expected remaining lifetime; \(D_{i,t}\) is the cost of deaths in the household, including an end-of-life and estate-settlement cost at the last death; and \(T_{i,t}\) is gifts and inheritances received. The model runs in wage-indexed dollars. Real wages grow at their 1969–2019 average of {{v:ws_g_cal}} a year in the calibration and on the 2026 Trustees Report’s intermediate path in the projections: {{v:ws_g_2535}} a year from 2025 to 2034 and {{v:ws_g_35}} thereafter . Results are in real 2024 dollars.

Returns. The return \(r_{i,t}\) follows the household’s portfolio. For each group, age band, and wealth band, the survey gives the dollars held in stocks, businesses, real estate, and safe assets, and we apply the long-run U.S. real returns estimated by Jordà and colleagues: 8.46% a year on equities including dividends, 4.19% in capital gains on businesses and 0.90% on real estate, and 1.92% on safe assets, with the year-to-year volatility they report , plus a persistent household-specific component. Where a cell holds fewer than {{v:ws_cell_min}} sampled families, as {{v:ws_fb_n}} of the {{v:ws_cells}} cells do, it pools the group’s age bands within its wealth band ({{v:ws_fb_pooled}} cells) or, where that is still too few, uses the other group’s portfolios at the same age ({{v:ws_fb_other}} cells: {{v:ws_fb_other_text}}). Black homeowners with mortgages earn 2.3 percentage points a year less on their homes, a gap Kermani and Wong trace to foreclosures and short sales rather than to slower appreciation .

Lives and families. Deaths follow the race-, sex-, and age-specific probabilities of Section 4, adjusted for wealth: within each group, death rates after 40 fall with household wealth in proportion to the hazard ratios of Machado and colleagues (0.80, 0.68, and 0.60 for the second, third, and top wealth quartiles relative to the lowest) , of which only a share is treated as causal. In each run that share is drawn between 0 and {{v:ws_lam_max}} of their logarithm: Swedish lottery wins rule out causal effects on 10-year mortality one sixth as large as the cross-sectional gradient , and inheritances in the Health and Retirement Study have no substantial effect on health . When its head reaches about 55, a household forms a successor household, standing for its children, whose income rank follows the intergenerational equations of Chetty and colleagues, in which a Black child’s expected rank is 12.6 percentiles below that of a White child whose parents had the same income . Parents with substantial wealth give a gift when the successor forms, and education debt falls with parents’ wealth. When the last member of a household dies, its estate, net of an end-of-life and settlement cost and of a share that leaves the households modeled, is split equally among the parents’ living children, whose number is drawn from the survey’s distribution by race (a mean of {{v:ws_kids_b}} for Black and {{v:ws_kids_w}} for White households aged 55 to 74); the successor takes one share and households of the same race and generation, standing for its siblings, take the others.

Calibration. {{v:ws_npar}} parameters, chiefly the saving schedule, the drawdown of wealth in retirement, and the size of gifts and estates, are chosen so that under unchanged conditions the model’s balanced-growth steady state reproduces median and mean net worth by age for both groups in the 2022 survey, and the share of households that have ever received $10,000 or more in inheritances and gifts, by age, measured as the survey asks it (Table {{n:tab-wcal}}). In the headline specification each group’s base saving rate is fitted separately: {{v:ws_s0b}} of income at the median income rank for Black households and {{v:ws_s0w}} for White households. The lower Black rate goes in the direction of the survey, in which Black households under 65 were less likely than White households to have spent less than their income in {{v:ws_saved_q}} of five income quintiles, but it is a residual that absorbs whatever the model leaves out, not a measured saving rate: version 2.0, with a different treatment of inheritances and no wage growth, fitted a higher Black rate ({{v:v20_s0b}} against {{v:v20_s0w}}), and in panel data racial differences in saving rates are not significant once income is controlled . We therefore also report a sensitivity with one saving schedule for both groups ({{v:ws_m_s0}} at the median rank, {{v:ws_m_npar}} parameters). It fits the survey worse (a loss of {{v:ws_m_loss}} against {{v:ws_loss}}, with one parameter fewer) and its backcast fails more often (Section 5.2), which is why it is not the headline. The calibration implies consumption out of wealth of {{v:ws_cw}} a year, against the 3.2 cents per dollar estimated for stock wealth , a figure it did not target. Its largest problem is inheritances. Splitting estates among children brings receipt near the survey, {{v:ws_rec_m_b}} of Black and {{v:ws_rec_m_w}} of White households against {{v:ws_rec_t_b}} and {{v:ws_rec_t_w}} (version 2.0: {{v:v20_rec_b}} and {{v:v20_rec_w}}), but only with an end-of-life and settlement cost of up to {{v:ws_care}} at the last death and a further {{v:ws_leak}} of each remaining estate leaving the households modeled; White recipients receive a median of {{v:ws_cmed_m_w}} against {{v:ws_cmed_t_w}} in the survey. Real estates do not shrink that much; the wedge more plausibly stands in for transfers the survey does not record and for late-life spending the model lacks. The one-schedule calibration needs a smaller wedge ({{v:ws_m_care}} and {{v:ws_m_leak}}) but overstates receipt ({{v:ws_m_rec_b}} and {{v:ws_m_rec_w}}). The headline calibration understates White mean wealth in {{v:ws_cal_wmean_low}} of {{v:ws_cal_nbands}} age bands and overstates the Black median by more than a tenth in {{v:ws_cal_bmed_high}}, at 65–74 by a factor of {{v:ws_cal_b6574}}. In its steady state the ratio of median wealth over all ages is {{v:ws_ss}}, against {{v:ws_svy_ratio}} in the survey ({{v:ws_m_ss}} with one saving schedule): the calibration targets wealth by age within each group, not the overall ratio, and does not reproduce the gap it starts from.

Uncertainty. Each path is simulated {{v:ws_runs}} times with {{v:ws_nh}} households, half Black and half White, reweighted to the population. Each run draws the long-run asset returns, the housing gap, the causal share of the wealth–mortality link, the cost of debt, and the Trustees’ low-cost, intermediate, or high-cost mortality from ranges around their estimates, and a different sequence of market returns. Markets, each group’s households, and each group’s heirs draw from separate random streams, so paths are compared run by run and a lever that acts only on Black households leaves every White household’s path unchanged. For each year we report the median across runs, a pointwise median rather than a single run, and the 5th to 95th percentiles across runs. These ranges contain the drawn parameters and Monte Carlo variation only; the calibrated parameters and the survey targets are held fixed, so they are not confidence intervals. The Monte Carlo standard error of a 2100 median ratio is at most {{v:ws_mcse_max}} in the headline specification and {{v:ws_m_mcse_max}} with one saving schedule. Where we date a closing in the median run, it is the run whose 2100 ratio is the median. In all, the analysis simulates about {{v:ws_total}} households.

Validation against earlier surveys

Fit to the targets is not validation, so we ran the model forward from the households of the 1989 and 1992 surveys through the realized conditions of each later year (S&P 500, house-price, and Treasury-bill returns, real wage growth, and historical death rates), with the calibration unchanged, and compared it with each later survey (Table {{n:tab-backcast}}). The pass criteria were written down before any run, with the one-schedule calibration in mind; they named 1989 as the start and 1992 as a check, but noted in advance that the 1989 survey’s small Black sample made 1992 the more reliable test, and we treat it as the primary one. The headline specification was chosen after both backcasts had been seen, so for that choice the backcast is evidence of fit, not a test made in advance.

From 1992 the headline specification keeps the ratio within 0.03 of the survey at all {{v:ws_bc_free_pass}} waves tested. Its 2022 Black median, {{v:ws_bc_h_b22_m}}, is {{v:ws_bc_h_b22_ok}} the survey’s 95% interval around {{v:ws_bc_b22_s}}, but Black median wealth grows by a factor of {{v:ws_bc_h_bmed_g}} against {{v:ws_bc_bmed_gs}} in the survey, {{v:ws_bc_h_bmed_ok}} the criterion. With one saving schedule for both groups, the model keeps the ratio within 0.03 at only {{v:ws_bc_pass}} of {{v:ws_bc_tested}} waves: it follows the ratio until 2004 but overstates it by {{v:ws_bc_over}} {{v:ws_bc_over_where}}, missing the depth of the Great Recession’s losses for Black households, although its Black median grows by a factor of {{v:ws_bc_bmed_g}}. Both specifications fail on White wealth: from 1992 the White median grows by a factor of {{v:ws_bc_h_wmed_g}} in the headline specification and {{v:ws_bc_wmed_g}} with one saving schedule, against {{v:ws_bc_wmed_gs}} in the survey, and the White mean by {{v:ws_bc_h_wmean_g}} and {{v:ws_bc_wmean_g}} against {{v:ws_bc_wmean_gs}}. The model gives all households one rate of consumption out of wealth and holds the income distribution at its starting shape, so large fortunes do not compound at realized returns and the rise in top incomes is missing. It therefore probably understates the growth of White wealth and overstates the ratio under current conditions. From 1989, run only with one saving schedule, the model passes at only {{v:ws_bc89_pass}} of {{v:ws_bc89_tested}} waves: the 1989 survey’s Black sample is {{v:ws_bc89_bfam}} families, and its ratio of {{v:ws_bc89_r}} becomes {{v:ws_bc89_r92}} in 1992, a jump no model of accumulation reproduces. Receipt of inheritances in the backcasts’ 2022 is low by construction ({{v:ws_bc_rec_b}} and {{v:ws_bc_rec_w}} with one saving schedule), because households start with no record of past receipt.

Under current conditions, the gap does not close

If today’s conditions persist, the ratio of median Black to median White household wealth, {{v:ws_base_2024}} in the model’s 2024, is {{v:ws_base_2050}} in 2050, {{v:ws_base_2075}} in 2075, and {{v:ws_base_2100}} in 2100, with 90% of runs between {{v:ws_base_2100_lo}} and {{v:ws_base_2100_hi}} (Figure {{n:fig-gap}}). This path is largely the model converging toward its own balanced-growth steady state, whose median ratio under unchanged conditions is {{v:ws_ss}} ({{v:ws_ss_lo}} to {{v:ws_ss_hi}} across {{v:ws_ss_seeds}} simulation seeds) at the calibration’s wage growth of {{v:ws_g_cal}} a year, above the survey’s {{v:ws_svy_ratio}}. With one saving schedule the steady state is {{v:ws_m_ss}} ({{v:ws_m_ss_lo}} to {{v:ws_m_ss_hi}}) and the 2100 ratio {{v:ws_m_base_2100}}. The level the ratio reaches by 2100 is therefore a property of the calibration, not a forecast that the gap narrows, and the model cannot attribute the movement to younger cohorts; because the backcasts understate the growth of White wealth, the path is probably optimistic. White wealth does not keep pace with wages: White household wealth is ${{v:ws_bw_2024}} trillion in 2024, ${{v:ws_bw_2050}} trillion in 2050, and ${{v:ws_bw_2075}} trillion by 2075, while real wages grow by a factor of {{v:ws_wage_2075}}, and Black household wealth goes from ${{v:ws_bb_2024}} trillion to ${{v:ws_bb_2075}} trillion. Relative to wages, the White median {{v:ws_wmed_wage_dir}} by {{v:ws_wmed_wage_2100}} by 2100 and the Black median {{v:ws_bmed_wage_dir}} by {{v:ws_bmed_wage_2100}}, so part of the movement in the ratio comes from White wealth falling relative to wages, which the backcasts, in which White wealth grows far too slowly, suggest is a weakness of the model rather than a forecast. Version 2.0, without wage growth, showed the ratio falling to {{v:v20_base_2050}} by 2050.

One lever at a time

Table {{n:tab-levers}} applies each lever alone. Single levers were run only with one saving schedule for both groups (the table’s lower block), so how each would differ in the headline specification is not known; for the one combination run under both, equal conditions, the headline gives {{v:ws_cond_2100}} by 2100 against {{v:ws_m_cond_2100}}. Earnings parity, in which Black workers reach the White income distribution by 2040 and Black children’s mobility matches White children’s, is the largest single lever: the median ratio rises to {{v:ws_earn_2050}} by 2050 and {{v:ws_earn_2100}} by 2100. It works slowly because higher earnings raise wealth only through saving and, over generations, through gifts and inheritances. Parity in mobility alone, which reaches only children, takes longer still ({{v:ws_mob_2100}} by 2100). Equal returns, which give Black households White portfolios and remove the housing-distress gap, raise the ratio to {{v:ws_ret_2100}} by 2100. Survival parity barely moves median wealth ({{v:ws_surv_2100}}); its value runs through Social Security, partners, and years of life (Section 4 and below) more than through net worth. Baby bonds of up to $50,000 for children born from 2025 reach {{v:ws_bb_2100}}. A one-time transfer of $250,000 to each Black household in 2030 lifts the ratio to {{v:ws_tr_2031}} at once, but under unchanged conditions it decays to {{v:ws_tr_2100}} by 2100, as Derenoncourt and colleagues found for transfers made without changing the conditions of accumulation . Structural models reach the same conclusion: in a general-equilibrium model, one-time transfers have only transitory effects unless the earnings gap closes , and in a model in which exclusion has left Black dynasties pessimistic about risky returns, transfers that close today’s average gap do not produce long-run convergence .

What closes the gap

Only combinations close the gap. In the headline specification, equal conditions (earnings and mobility parity by 2040, equal returns by 2035, and survival parity by 2050) raise the median ratio to {{v:ws_cond_2050}} by 2050, {{v:ws_cond_2075}} by 2075, and {{v:ws_cond_2100}} by 2100 (90% of runs between {{v:ws_cond_2100_lo}} and {{v:ws_cond_2100_hi}}) if Black households keep their lower saving and their support to relatives. The ratio is at least 0.9 in {{v:ws_cond_share}} of runs in 2100, and the median run does not stay at or above 0.9 by 2100. Equal conditions stop the forces that keep the gap open but leave the initial difference in wealth, and the difference in saving, to be worked off across generations. Equal saving is therefore a separate lever: if Black households’ base saving rate and their support to relatives also reach the White ones by 2035, the ratio reaches {{v:ws_csav_2050}} by 2050, {{v:ws_csav_2075}} by 2075, and {{v:ws_csav_2100}} by 2100, at least 0.9 in {{v:ws_csav_share}} of runs; half the runs stay at or above 0.9 from {{v:ws_csav_year}} or earlier, and the median run does so from {{v:ws_medrun_csav}}. Because the saving gap is a fitted residual, not a measured behavior (Section 5.1), what would move it, and whether it reflects saving at all rather than something the model leaves out, is unknown. With one saving schedule for both groups, equal conditions reach {{v:ws_m_cond_2100}} by 2100 ({{v:ws_m_cond_share}} of runs at or above 0.9), and {{v:ws_m_csup_2100}} with equal family support as well; without the end-of-life cost in that calibration, the latter reach {{v:ws_nocare_cond_2100}}. With the one-time transfer added to equal conditions and equal saving, half the runs close the gap from {{v:ws_condtr_year}}, when the transfer is paid, and the median run from {{v:ws_medrun_condtr}}; the gap stays closed, and {{v:ws_condtr_share}} of runs are closed in 2100 (with one saving schedule, half the runs close from {{v:ws_m_condtr_year}}). To find the smallest steps that suffice, we screened the combinations of levers and reran the most promising at full size, {{v:ws_hh_per}} households each (Table {{n:tab-bundles}}). The screen was run only with one saving schedule for both groups, so whether the same combinations suffice in the headline specification is not known. {{v:ws_bund_text}} Without a transfer, earnings parity and equal returns close the gap by 2100 only at the edge of the criterion, in {{v:ws_notr_shares}} of full-size runs. Earlier parity raises the ratio sooner ({{v:ws_notr_e2050}} in 2050 with parity by {{v:ws_notr_early}}, against {{v:ws_notr_l2050}} with parity by {{v:ws_notr_late}}) but ends lower in 2100 ({{v:ws_notr_e2100}} against {{v:ws_notr_l2100}}), so these shares do not rise with earlier parity. The criterion has no upper bound, and the closing combinations overshoot parity, with median ratios of {{v:ws_bund_med_range}}.

Table {{n:tab-shapley}} divides the change in the median ratio among four levers by their Shapley values, each lever’s added effect averaged over every order in which the levers could be added, computed run by run from the screening grid, which was run with one saving schedule for both groups. By 2100, earnings parity contributes {{v:ws_sh_inc_2100}} and equal returns {{v:ws_sh_ret_2100}}, although equal returns alone add only {{v:ws_sha_ret_2100}}. That gap does not mean the levers amplify each other. Each lever raises Black wealth by roughly a constant factor, so a lever moves the ratio more once another lever has raised Black wealth: on the ratio, the four levers together add {{v:ws_lvladd_range}} times the sum of their effects alone between 2050 and 2100, but on the logarithm of the ratio they add {{v:ws_logadd_range}} times that sum, close to additive. The independent re-implementation of Section 5.8 finds the same in log wealth. Survival parity contributes {{v:ws_sh_surv_2100}}. The transfer contributes {{v:ws_sh_tr_2050}} in 2050 but {{v:ws_sh_tr_2100}} by 2100. That earnings dominate agrees with the general-equilibrium model of Aliprantis, Carroll, and Young and with Barsky and colleagues’ estimate that differences in earnings account for about two thirds of the Black–White gap in mean wealth .

Three implications follow within the model. Transfers last only when the forces that widen the gap stop: unequal pay and opportunity, and unequal access to diversified assets that do not have to be sold in distress. The initial difference in wealth is worked off slowly through children’s endowments, which do not close the gap this century, or quickly through a capital transfer of $250,000 to $500,000 per Black household, about ${{v:ws_tr_cost_250}} trillion to ${{v:ws_tr_cost_500}} trillion in all. And closing the survival gap barely moves the wealth ratio but accounts for most of the gains in years of life.

What closing the gap would add

Table {{n:tab-gains}} compares current and equal conditions in the headline specification, with Black households keeping their lower saving. By 2050, equal conditions would add about ${{v:ws_gain_w2050}} trillion to Black household wealth, and ${{v:ws_gain_w2075}} trillion by 2075 (${{v:ws_gain_sav_w2075}} trillion with equal saving as well; ${{v:ws_m_gain_w2075}} trillion with one saving schedule and equal family support), in real 2024 dollars on a path where real wages grow by a factor of {{v:ws_wage_2050}} between 2024 and 2050 (version 2.0, without wage growth: ${{v:v20_gain_w2075}} trillion by 2075). Black household income before 67 would be ${{v:ws_gain_inc2050}} trillion a year higher by 2050, or ${{v:ws_gain_inc2050_w24}} trillion in 2024 wage terms, {{v:ws_gain_gdp2050}} of 2024 gross domestic product . How much of that is added output rather than redistribution depends on effects the model does not include. The added wealth alone would support about ${{v:ws_gain_cons2050}} billion a year in consumer spending at the measured propensity to consume out of wealth . Through survival parity and the wealth–mortality link, life expectancy at 40 would rise by 2075 from {{v:ws_e40m_base}} to {{v:ws_e40m_cond}} years for Black men and from {{v:ws_e40w_base}} to {{v:ws_e40w_cond}} years for Black women. Survival parity, which is imposed rather than produced by any lever, accounts for most of that gain. How much wealth adds depends on the share of the wealth–mortality gradient treated as causal: in runs with one saving schedule and that share fixed at zero, one sixth, and one half, earnings parity alone adds {{v:ws_lam_earn_0}}, {{v:ws_lam_earn_16}}, and {{v:ws_lam_earn_50}} years to Black men’s life expectancy at 40 (version 2.0, with a share between a quarter and three quarters: {{v:v20_e40m_earn}}), survival parity alone adds {{v:ws_lam_surv_0}}, {{v:ws_lam_surv_16}}, and {{v:ws_lam_surv_50}}, and equal conditions with equal family support give {{v:ws_lam_cond_0}}, {{v:ws_lam_cond_16}}, and {{v:ws_lam_cond_50}} years. The death rate of Black adults aged 25 to 64 would fall from {{v:ws_dr_base}} to {{v:ws_dr_cond}} per 100,000, the share of Black couples aged 30 who lose a partner before 67 from {{v:ws_partner_base}} to {{v:ws_partner_cond}}, and the share of Black children who lose both parents by 45 from {{v:ws_orph_base}} to {{v:ws_orph_cond}}.

Where the gains would land

The microsimulation is national, so it does not say where its gains would arise. To show where they would be largest, we apportion them to the 50 states and the District of Columbia. For each state, the 2020–2024 American Community Survey Public Use Microdata Sample gives the number of households headed by a non-Hispanic Black person and the mean income of non-Hispanic Black and of non-Hispanic White households, the groups of the national model . From these we compute, for each state, the income its Black households would add if their mean income reached that of White households in the same state, and divide the national gains among the states in proportion to that shortfall. The states’ shares do not depend on the specification; the dollar amounts below apportion the gains of equal conditions with equal family support under one saving schedule for both groups (${{v:ws_m_gain_w2075}} trillion by 2075), the national figure the state view was built on, rather than the headline’s ${{v:ws_gain_w2075}} trillion. This is an apportionment, not a state-level simulation: it assumes that the gains would arise where today’s shortfall is, and it ignores migration and differences between states in prices, growth, and returns.

The gains would follow the Black population. The Census Bureau’s South region, home to {{v:st_south_hh}} of Black households, would receive {{v:st_south_share}} of the national gain, and ten states would receive {{v:st_top10_share}} (Table {{n:tab-states}}). {{v:st_top3}} lead; {{v:st_top1}} alone would gain about ${{v:st_top1_w2075}} trillion in Black household wealth by 2075. Relative to the size of their economies, the gains would be largest in {{v:st_gdp_top2}}, where the added income in 2050, in 2024 wage terms, would equal about {{v:st_gdp_top2_pct}} of their 2024 gross domestic product , against {{v:ws_m_gain_gdp2050}} for the nation. Apportioning by each state’s share of Black households instead gives almost the same ranking (Spearman correlation {{v:st_rho}}); weighting by the income gap raises states where that gap is wider per household, such as {{v:st_wider}}. Counting Black householders of any ethnicity, as the published summary tables do and version 2.0 of this paper did, moves no state’s share by more than {{v:st_bt_maxpp}} percentage points (the largest change is in {{v:st_bt_max_state}}; Spearman correlation {{v:st_rho_bt}}). The survey’s 80 replicate weights put the 90% margin of error of each of the ten largest shares at {{v:st_top10_moe_max}} percentage points or less. Estimates for {{v:st_low_text}} ({{v:st_low_names}}) are imprecise, and the replication data flag them.

What the microsimulation adds. The wealth gap is not a fixed sum to be paid off once. It is a flow, sustained by unequal earnings, unequal returns, and unequal lives acting on an unequal starting point. The model puts dates and dollars on what it would take to stop the flow and close the stock, and on what closing it would return. Its numbers depend on its assumptions, above all on whether the two groups save alike at equal income, and all of them are in the replication files.

An independent replication

Before this version, we built a second model of the same question from the same primary data, with separate code written independently of the first and without reading the first model’s output until the second had been specified, calibrated, and run, and compared the two using version 2.0 of this paper. It is a re-implementation of the question, not of the code: an annual accounting model with cells by race, age, and income rank, seven calibrated parameters, {{v:xc_draws}} paired draws, and four levers defined differently from ours (survival parity; parity in the receipt of inheritances and in family help in emergencies; equal access to assets, including portfolios, leverage, and housing returns; and parity in high-cost interest, fees, and stress spending), with earnings parity as a reference path. Its inputs reproduce ours: {{v:xc_t1_match}} entries of Table {{n:tab-survival}} that it recomputed from the life tables agree to the printed precision, and it puts the age-composition channel at {{v:xc_comp_range}} of the log wealth gap, with medians and with means by age, against {{v:xc_comp_paper}} in Section 3. Its levers are additive in log wealth: all four together closed {{v:xc_amp_range}} times the sum of what each closed alone at horizons from 2035 to 2100 (5th to 95th percentiles across draws from {{v:xc_amp_lo}} to {{v:xc_amp_hi}}), the same near-additivity our grid shows on the logarithm of the ratio (Section 5.5). Combinations are needed because no single lever is large enough, not because levers amplify each other. Its calibrated saving residual has the sign opposite to version 2.0’s: Black households save less ({{v:xc_s0b}} against {{v:xc_s0w}} at the median rank), where version 2.0 found them saving more. That the sign follows the specification is why this version treats the saving gap as a fitted residual and equal saving as a separate lever. A one-time transfer fades in it too: {{v:xc_tr_amt}} per Black household in 2030 closes {{v:xc_tr_2035}} of the log gap by 2035 and {{v:xc_tr_2100}} by 2100 under unchanged conditions. Its equal-access lever is the largest after earnings (about {{v:xc_ret_2075}} of the log gap by 2075), and the lever on high-cost interest, fees, and stress spending is small: about {{v:xc_fee}} per Black household a year in 2050 and {{v:xc_fee_share}} of the log gap by 2075. Like ours, it finds that survival parity barely moves median wealth and that earnings parity is the largest lever (a ratio of {{v:xc_earn_2100}} by 2100). It was run against version 2.0, not this version; its outputs are summarized in the replication files.

Quality of life in the later years

Longevity counts years; people live in them. In 1980 Fries argued that if the onset of chronic illness could be postponed faster than death, illness would be compressed into a short period at the end of life . Whether that happens for a given person depends heavily on resources.

Healthy years. In harmonized data from England and the United States, the socioeconomic disadvantage in disability-free life expectancy was largest for wealth: at age 50, people in the poorest group could expect seven to nine fewer years without disability than those in the richest . Among Americans aged 54–64, those in the poorest wealth quintile had more than three times the ten-year risk of disability of those in the richest (48% versus 15%) . Racial gaps extend to healthy years: in life-table models built from 1990 census data, Black Americans lived the fewest years of the major racial groups and spent a high proportion of them with a chronic health problem .

Money in later life. Under the Census Bureau’s official measure, 9.8% of Americans aged 65 and older were poor in 2025. Under the Supplemental Poverty Measure, which subtracts out-of-pocket medical spending and accounts for taxes and benefits, the rate was 15.4%, the highest of any age group, and medical expenses alone raised the older-adult rate by 3.8 percentage points . Poverty in later life is unequal by race: under the official measure, 17.3% of Black adults aged 65 and older were poor in 2020–2022, more than twice the rate among older White adults, 7.7% . Social Security lifted 20.9 million people aged 65 and older out of poverty under the supplemental measure in 2025 , which is why the annuity channel in Section 3 matters.

Decisions in later life. Financial mistakes follow a U-shape over the life cycle; across ten kinds of credit decisions, they are least costly around age 53 and become more common with age thereafter . Financial trouble can also precede a dementia diagnosis: Medicare beneficiaries living alone who were later diagnosed with Alzheimer disease or a related dementia were slightly more likely than similar beneficiaries to miss credit payments as early as six years before diagnosis (7.7% versus 7.3%), and the difference grew after diagnosis . The early difference is small, so a missed payment says little about any one person; the finding argues for consent-based safeguards, such as a trusted contact, rather than for inferring anyone’s health from their spending.

Connection. Across 148 studies, people with stronger social relationships had a 50% greater likelihood of survival over the follow-up periods studied, an effect comparable to quitting smoking . Wealth buys more than care; it sustains participation in family and community life.

On this evidence, a good later life requires enough wealth to compress illness into the end of life rather than spread it across decades; protection from costly mistakes and exploitation as cognition changes; and the autonomy and connection that make long life worth having. All three are distributed unequally by wealth and therefore by race.

Stress spending: the everyday mechanism

Chronic stress shortens lives through the body. It may also drain wealth through behavior; one everyday route is spending that soothes a state rather than serves a goal. We use stress spending for purchases whose timing and content are better explained by the buyer’s momentary state (sadness, exhaustion, anxiety, depletion) than by their plans. This is a construct for research, not a validated measure.

Emotion carries over into prices

Incidental emotions, unrelated to the purchase at hand, change what people will pay. In experiments with real goods and money, induced sadness raised the prices people chose to pay and lowered the prices at which they would sell, reversing the usual endowment effect ; sad participants who were also self-focused spent more on the same item . Shopping can also repair mood. Making purchase decisions reduced residual sadness by restoring a sense of personal control , and unplanned treats bought to repair a bad mood were not, on average, followed by regret or guilt . A product that treated every comfort purchase as pathology would therefore be wrong as well as unkind. The target is the pattern that hurts: repeated, escalating, regretted spending.

The body sets the stage

In a laboratory study, one night without sleep moved healthy volunteers’ risky choices away from avoiding losses and toward pursuing gains, with matching changes in brain activity . Heart rate variability (HRV) is linked to activity in brain regions that regulate emotion and threat and is widely used as a marker of stress and regulatory capacity . Sleep debt, lower HRV, and higher resting heart rate, each measured against a person’s own baseline, are therefore plausible but unvalidated indicators of the moments when self-regulation is hardest: the studies above used laboratory sleep deprivation and brain imaging, and none of the work we cite shows that consumer-device measures of sleep or heart rate predict regretted spending.

Scarcity narrows attention

Prompting people to think about a costly financial problem reduced cognitive performance among lower-income participants but not among higher-income ones, and the same sugarcane farmers performed worse before harvest, when poor, than after it . Scarcity focuses attention on the present problem and leads people to over-borrow against the future . Financial stress can impair the very capacities needed to escape it. Some of this literature has not held up, however: an audit that repeated 20 studies of the psychological consequences of scarcity found considerable variation in how well they replicated, with some strong successes and some clear failures .

Mental health and money move together

In a self-selected online survey of 5,413 people in the United Kingdom with lived experience of mental health problems, 93% said they had spent more than usual when unwell, 92% found financial decisions harder, and 59% had taken out a loan they otherwise would not have . A meta-analysis found that people with unsecured debt had about three times the odds of a mental disorder (odds ratio 3.24), although the direction of causation is unclear , and among 8,400 young American adults, high debt relative to assets was associated with higher perceived stress and depression, worse self-rated health, and higher diastolic blood pressure .

Products can make it worse

After states legalized online sports betting, betting spread quickly and savings fell as risky bets crowded out investment; among financially constrained households, credit card debt and overdrafts rose and available credit fell . Always-on access to speculative or compulsive spending is not neutral for households under strain.

The loop, and where to break it

Together these findings describe a loop (Figure {{n:fig-loop}}): adversity raises stress; stress, sleep loss, and low mood change what people buy; unplanned spending and debt raise financial strain; and financial strain is itself a chronic stressor that wears down the body. The loop can be interrupted at several points. Income support, debt relief, and health care act on its largest drivers. A smaller point of intervention is the moment just before the next purchase, when a little context, a little friction, or a commitment made in a calmer moment might change what happens next. Defaults , precommitment to future behavior , and self-chosen commitment devices have changed financial behavior in other settings; whether the same tools, timed by context, can interrupt stress spending is untested.

Other wealth gaps where the same forces may operate

The mechanisms in Sections 2–7 are not specific to race. Wherever savings must stretch over more years than they can cover, wherever stress is chronic, and wherever products profit from moments of weakness, wealth gaps can open. This section summarizes published evidence on eight such gaps; Table {{n:tab-gaps}} lists the evidence for each, with its year and measure, and the mechanism behind it. It does not estimate how much each mechanism contributes to any gap (Section 9).

Women: more years, fewer dollars. Women live longer. In 2024, life expectancy at 65 was {{v:e65_f}} years for women and {{v:e65_m}} for men, and a woman alive at 22 can expect {{v:ret_f}} years of life at 67 and older against {{v:ret_m}} for a man, {{v:ret_f_pct}} more . Those years must be funded from smaller balances. Among unmarried householders, women had less median wealth than men under 35, at 35–54, and at 65 and over; unmarried women under 35 held a median of $8,040, 30.9% of their male counterparts’ wealth . Older women are more often poor: in 2020–2022, 11.0% of women aged 65 and over were below the official poverty line, against 8.5% of men . The partner channel in Section 3 adds to the burden. If both partners are 30 and their risks are independent, the 2024 life tables give the woman a {{v:outlive}} chance of outliving the man ({{v:outlive_gap2}} if he is two years older), so the longest stretch of later life is most often funded by a woman alone.

Hispanic Americans: long lives, thin buffers. Hispanic Americans live longer than non-Hispanic White Americans ({{v:e0_h}} years at birth against {{v:e0_w}} in 2024) , a pattern described in 1986 as an epidemiologic paradox because it coexists with lower incomes . Their wealth does not match their longevity. In 2022 the typical Hispanic family held $61,600, about 22% of the typical White family’s $285,000 ; 15.5% of Hispanic households had zero or negative wealth ; and among middle-aged families, 28% of Hispanic families held a retirement account, against 65% of White families . A Hispanic 22-year-old can expect {{v:ret_h}} years of life at 67 and older, {{v:ret_h_pct}} more than a White 22-year-old, to be financed with a fraction of the savings. Older Hispanic adults are poor at more than twice the rate of older White adults (17.4% against 7.7%) . Here longevity is an advantage that the wealth gap turns into a risk: outliving one’s savings.

American Indian and Alaska Native people: the shortest lives. In 2024, American Indian and Alaska Native people had the lowest life expectancy of any racial or ethnic group in the life tables, {{v:e0_ai}} years, and only {{v:p65_ai}} of a birth cohort survives to 65 . Older American Indian and Alaska Native adults are poor at the same rate as older Hispanic adults, 17.4% . The survival channels in Section 3 are larger for this population: under 2024 death rates, an American Indian or Alaska Native 22-year-old would have a {{v:p2267_ai}} chance of reaching 67 (White: {{v:p2267_w}}) and could expect {{v:ret_ai}} years of life after 67, {{v:ret_ai_pct}} fewer than a White 22-year-old. Old-age mortality for this population is modeled in the life tables, which adds uncertainty . Yet the Census Bureau’s wealth report and the Survey of Consumer Finances both fold American Indian and Alaska Native households into a combined “other” category , so the wealth gap is poorly measured. A gap that is not measured is hard to close; we flag it rather than estimate it.

Disability and mental health. In 2023, 25.3% of working-age people with disabilities lived in poverty, against 11.4% of those without . Poor mental health and money problems are closely linked: in the self-selected U.K. survey described in Section 7, 93% of people with lived experience of mental health problems said they had spent more than usual when unwell , and unsecured debt is associated with about three times the odds of a mental disorder . Money can be harder to manage when one is unwell, and the Money and Mental Health Policy Institute has called for ways to help people avoid what it calls “financial self-harm” .

Gambling harm. Where sports betting has been legalized, household finances have measurably worsened. In credit-report data on roughly 7 million consumers, average credit scores fell by about 0.7 points where any legal sports betting became available and by about 12 points where online or mobile betting was added in states that already allowed retail betting, alongside increases in bankruptcy filings, debt sent to collections, credit card delinquencies, and auto loan delinquencies ; among financially constrained households, savings fell and credit card debt and overdrafts rose .

Penalty fees: the scarcity tax. Fees concentrate on households with the least slack. In data from large banks, the 9% of accounts with more than ten overdraft or nonsufficient-funds transactions a year paid 79% of all such fees . Combined with scarcity’s toll on attention , small, avoidable shortfalls become a steady drain on the households with the least wealth.

Values-based exclusion. Some households avoid interest-bearing products, or particular industries, for religious or ethical reasons. Evidence on how much this keeps people out of the financial system comes from outside the United States and is dated: in 2011 survey data, religious reasons were among the barriers people reported for not having an account, most often in the Middle East and North Africa and in South Asia, and Demirgüç-Kunt and Klapper estimated that products compatible with religious beliefs could raise account ownership there by up to 10 and 5 percentage points respectively, while cautioning that such analysis cannot support causal statements . When the only savings products conflict with a household’s values, saving may happen in cash or not at all.

Age. Families headed by someone under 35 held a median of $39,000 in 2022 , and costly financial mistakes are more common among both younger and older adults than in middle age . In later life, missed payments become slightly more common years before a diagnosis of dementia .

Across these gaps the pattern matches Section 3. Longer or shorter lives change how far wealth must stretch, chronic stress drains it, and products that profit from hard moments can widen gaps fastest among those with the least room for error. Any product aimed at these mechanisms should be evaluated for harm as well as benefit in exactly these households.

Threats to validity

Table {{n:tab-threats}} lists the main threats to the results of Sections 3 to 5, what this version did about each, what changed as a result, and what remains. Four matter most for the headline. First, whether Black and White households save alike at equal income: the headline specification, which fits the survey and its history better, has Black households saving less, and under equal conditions it gives {{v:ws_cond_2100}} by 2100 unless saving also equalizes ({{v:ws_csav_2100}}); with one saving schedule for both groups the same conditions give {{v:ws_m_cond_2100}}. The saving gap is a fitted residual, and the headline was chosen after the backcasts were seen. Second, the model fits the survey’s inheritance receipt only by removing most estate wealth. Third, it reproduces the history of the Black median from 1992 better than that of White wealth, whose growth it understates, so the current-conditions path is probably optimistic. Fourth, that path largely converges toward the calibration’s steady state ({{v:ws_ss}}), which is above the survey’s ratio, so its level in 2100 is a property of the calibration. The single levers, the combinations, the Shapley values, the runs with fixed causal shares, and the state view were run only with one saving schedule. The ranges reported for the microsimulation contain the drawn parameters and Monte Carlo variation, not the uncertainty of the calibration or of the survey samples, and the calibration may not be fully converged. The reviews in Sections 2, 6, 7, and 8 rest largely on observational studies, and several of the sources they cite were checked against their abstracts rather than their full texts; the replication files list them.

Conclusion

Years and dollars are linked in both directions: wealth buys time, and time builds wealth. In the 2024 life tables, most of the Black–White gap in life expectancy arises from deaths before 65. The channels through which that survival gap can be measured directly in net worth, the age structure of the adult population and the length of working lives, account for only about {{v:comp_share}} and {{v:work_share}} of the log wealth gap. Through the channels net worth does not record, the survival gap is worth more. In an open projection model anchored to the Social Security Trustees’ assumptions, shorter lives cost a Black worker entering employment today {{v:pj_pen_f}} to {{v:pj_pen_m}} of expected Social Security wealth in present value, counting disability and family benefits, add {{v:pj_lost_pen}} to a Black couple’s expected loss of earnings to a partner’s death, and make it nearly twice as likely that a Black child loses both parents before 45. Survivor benefits replace about {{v:pj_surv_all_offset}} of that added loss, and bequest gaps reflect parents’ wealth rather than their lifespans. Falling mortality shrinks these penalties slowly. If Black death rates converged to White rates by 2050, the Social Security penalty would nearly vanish, but about half the partner-loss penalty would remain, because it accrues in working years before 2050. A microsimulation of household wealth shows what closing the gap would take. Under current conditions it does not close; the ratio of median wealth is {{v:ws_base_2100}} in 2100, close to the calibration’s own steady state, and probably lower in reality, since the model understates the past growth of White wealth. No single lever closes it by 2100. In the specification that fits the survey best, Black households save less at equal income, and equal earnings, mobility, returns, and survival raise the across-run median ratio to {{v:ws_cond_2100}} by 2100, at least 0.9 in {{v:ws_cond_share}} of runs. Equal saving is a separate lever: with it the ratio reaches {{v:ws_csav_2100}}. With one saving schedule for both groups, the same conditions reach {{v:ws_m_cond_2100}}. A capital transfer closes the gap from {{v:ws_condtr_year}} once those conditions and equal saving hold; alone, it fades. Equal conditions would add about ${{v:ws_gain_w2075}} trillion to Black household wealth by 2075 and, mostly through survival parity, years of life. An independent re-implementation with separate code reproduces the inputs and finds the levers additive in log wealth: combinations are needed because no single lever is large enough, not because the levers amplify each other. The racial wealth gap remains primarily a product of history and of continuing differences in income, inheritance, homeownership, and access to credit; in the model, it persists until those differences end.

Why we exist

The Understanding Company exists because we think one moment that matters for financial health is invisible to the institutions that hold people’s money. A bank sees an amount, a merchant, and a time. It does not see that a purchase came at 11:48 p.m. after a night with two hours less sleep than usual, that it is the third order of the evening, or that the last order from this merchant went back. A health app sees the sleep and the heart rate but not the purchase. Each holds half of the story, and neither acts on it.

We believe financial health is a health behavior and should be designed like one. That means treating the moment before a purchase as a moment of care: understanding why someone is spending, not only what; offering friction the person chose in advance; and protecting essential purchases. It also means taking seriously the evidence in Sections 2–8: chronic stress and thin buffers are associated with shorter lives, wealth and years compound together, and the burden of both falls hardest on households with the least room for error. Vueni does not target any racial, ethnic, gender, or disability group, and none of these characteristics is an input to any of its models.

Our product, Vueni, is a card and account built on seven commitments:

1. Understanding, not judgment. Every signal Vueni acts on is shown to the member in plain language (“Short sleep · 2h 12m below your usual”). An unexplained score reads as judgment; an explained one reads as understanding.
2. Act before, not after. Vueni’s interventions take effect ahead of the next purchase, while a choice can still be made, rather than arriving as a lecture after the money is gone.
3. Essentials protected. Groceries, pharmacy, fuel, transit, utilities, insurance, and medical care are never flagged, and Calm Mode’s limit is set above the member’s typical large essential purchase (Section 13.5), so that most essential purchases go through; any declined essential purchase is counted as a harm (Section 14).
4. The member is in charge. Calm Mode is off, suggest, or auto, chosen by the member, and the default is suggest. A guardrail the member sets has a cooldown before it can be switched off: a commitment chosen in a calm moment for a hard one.
5. Privacy by architecture. Health data is reduced to three numbers on the phone, raw samples never leave it, and health data is never used for credit, underwriting, or advertising.
6. Values-aware, not values-imposing. Values mode lets members choose guardrails, interest-free savings, and a giving rule that fit their beliefs. Vueni does not certify any product as meeting a religious standard.
7. Measure outcomes and publish them. Section 14 sets out what we plan to measure and report, including results that do not flatter us.

Table {{n:tab-mech}} lists the mechanisms from Part I that these features are aimed at. The features are matched to mechanisms, not to groups.

We are not a bank: Vueni’s accounts, cards, and money movement are provided through Whop’s financial infrastructure and its partners. Nor do we claim that a card can close a racial wealth gap built over centuries. Our hypothesis is narrower: that a product that understands stress, buffers shocks, and helps people keep what they earn can improve everyday quantities, such as unplanned debt, regretted purchases, and savings that survive a bad month. Section 14 describes how we plan to test it.

What the model asks of the product

Section 5 shows which levers close the wealth gap and what each is worth. Most belong to employers, schools, health systems, and governments: earnings, mobility, survival, transfers, and endowments. A consumer financial product can reach four, all of which matter most for households with thin buffers: the share of income saved; the share lost to high-cost interest , fees, and stress spending; whether a shock forces the distressed home sale behind the Black–White gap in housing returns ; and whether savings earn diversified returns. Vueni is built around these four (Table {{n:tab-mech}}). This section uses the microsimulation to set targets for them.

Effects we assume and must test

We have not measured Vueni’s effects (Section 14). For planning we assume that an active member raises saving by 2 percentage points of income, halves high-cost interest, cuts stress spending by half a percent of income, halves the housing-distress gap through emergency buffers, and earns half a percentage point more on positive wealth, and that 90% of members remain active each year. The saving assumption is small next to the effects of automatic enrollment and commitment devices in workplace and field settings ; the others are hypotheses that the evaluation in Section 14 is designed to test. Eligibility does not depend on race: households under 67 whose net worth is below the national median for their age.

Reach targets

Table {{n:tab-reach}} and Figure {{n:fig-reach}} show what these assumptions imply. Reaching half of eligible households, about {{v:ws_prod_adopt2030}} million households in 2030, {{v:ws_prod_black2030}} million of them Black, would add about {{v:ws_prod_med2050}} to the wealth of the median Black household by 2050 (90% interval, {{v:ws_prod_med2050_ci}}) and raise the median ratio by {{v:ws_prod_dratio2050}}; reaching every eligible household would raise it by {{v:ws_prod_full_dratio2050}}. With equal conditions and equal family support (one saving schedule), the same half reach would add {{v:ws_prod_cond_med2050}} to the median Black household and {{v:ws_prod_cond_dratio2050}} to the ratio. Among the levers of Section 5 the product is small: its Shapley value for the 2100 median ratio, computed with the six levers of the screening grid, is {{v:ws_sh_prod_2100}}, and added to equal conditions with equal family support, under one saving schedule for both groups, it moves the 2100 ratio from {{v:ws_m_csup_2100}} to {{v:ws_condprod_2100}}, within simulation error. The product cannot close the gap; it can narrow it and speed the work of the levers that can. Our targets follow from the model: reach half of eligible households by 2030, keep nine in ten active each year, and demonstrate the assumed effects in a randomized evaluation before claiming them.

The harness and the models

This section describes the Understanding engine as it exists in our codebase on October 1, 2026. The engine is pure, deterministic TypeScript: given the same member snapshot, it returns the same scores, matches, cycles, and interventions. It is covered by 45 unit tests, and every spending-related number in the app screens on our website (scores, cycles, forecasts, suggestions) is computed by the engine on a synthetic member rather than typed by hand. The iOS app that will supply health signals is not yet built, the engine has not been evaluated with members, and every weight below was set by hand. We report them so that they can be criticized. The engine does not use race, ethnicity, gender, or disability as inputs, and it does not attempt to infer any health condition, disability, or cognitive decline. Some of its inputs, such as health signals, time of day, and merchant category, are correlated with those characteristics and with shift work and income, so Section 14.2 plans an audit for disparate impact.

Constraints that shaped the design

Three facts about the platform shaped the engine. First, the card platform reports transactions after authorization and offers no real-time approve-or-decline hook, so the engine cannot stop a purchase in flight; it acts ahead of the next one by adjusting the card’s per-purchase limit, pausing the card briefly, or scheduling a reminder. Second, the card program declines some merchant categories at authorization by default, including the gambling codes; gambling-adjacent merchants that code themselves differently, such as sweepstakes casinos and prediction markets, can be detected only after the charge. Third, health data must stay on the phone. The iOS app will compute three deltas against the member’s own 14-day baseline (minutes of sleep below baseline, percent change in overnight HRV, and percent change in resting heart rate) and send only those; the server route accepts nothing else.

The context graph and receipt matching

A card transaction is the anchor; receipts, shipments, and health signals attach to it (Figure {{n:fig-engine}}). Receipts arrive by email forwarding. A rule-based parser extracts the total, order number, items, return window, and tracking numbers; emails it cannot parse are reserved for a language-model extraction pass with the same output schema. A receipt is matched to the transaction it paid for by a weighted confidence

where \(s_{\text{amount}}\) is 1 for an exact match to the cent, declines linearly to 0.4 at a 3% difference (absorbing tax, tips, and split shipments), and is 0 beyond; \(s_{\text{merchant}}\) is a bigram similarity after stripping payment-processor prefixes; and \(s_{\text{time}}\) declines linearly to 0 at three days. A match requires \(c \ge 0.62\); otherwise the engine reports no match rather than guessing.

The stress score

For each purchase \(t\), the version-0 stress score is additive:

where \(\mathcal{K}_t\) is the set of factors present, \(p_k\) are the points in Table {{n:tab-factors}}, and \(\mathcal{C}\) is the set of contextual factors (time of day, sleep debt, HRV, and resting heart rate). The bucket weight \(\beta\) is 0 for essential merchant categories, 0.5 for uncategorized ones, and 1 for discretionary ones, and the size weight \(\sigma\) is 0.5 for purchases under $15 and 1 otherwise. Context therefore never adds points to an essential purchase and counts for less on a small coffee after a bad night. An essential purchase can still earn the non-contextual points for an unusually large amount or for a merchant with past returns, at most 22 points, which is below the elevated threshold, so essentials are never flagged. Health factors apply only when the latest signal was observed no more than 18 hours before the purchase or up to one hour after it. Scores of 35 and above are elevated; 60 and above, high. The model is deliberately simple: every point traces to a reason the member can see, so the explanation is the model. It is a heuristic built from the evidence in Section 7, not a validated measure of stress spending.

Cycles

Stress spending is rarely a single purchase. The engine names three recurring shapes back to the member. A regret loop is two or more stressed purchases (score of at least 35) refunded within 45 days. A recurring window is three or more stressed purchases on the same weekday and in the same part of the day (mornings, afternoons, evenings, or nights, with purchases after midnight counted with the previous evening) across at least two weeks. A binge is three or more stressed purchases within six hours.

Interventions and Smart Nudges

Interventions follow a ladder that matches the strength of the response to the strength of the evidence (Table {{n:tab-interventions}}). When the most recent purchase, made within the past two hours, is high-stress, Calm Mode proposes, or in auto mode applies, a per-purchase limit until 8 a.m. local time; if the member has enabled it and the purchase is part of a binge, it proposes a three-hour pause instead. The limit is

where \(Q_{0.9}(E)\) is the member’s essential purchase at the 90th-percentile rank (with the member’s n essential purchases sorted from smallest, the one in position ⌊0.9n⌋ + 1, which is the largest when n ≤ 10) and \(F\) is a floor of $50. An elevated purchase triggers only a Smart Nudge, a short in-app or push message such as “Heavier day than usual. Want to sleep on the next one?” When a recurring window is known, Calm Mode is offered an hour before it opens. Stressed purchases with a stated return window generate a reminder three days before it closes. A week without a high-stress purchase earns a week of 1% cashback. Every limit and pause reverts automatically.

Forecast

A cash-flow forecast identifies recurring flows (paychecks, rent, subscriptions) as three or more flows with the same counterparty at a weekly, biweekly, or monthly cadence whose amounts vary by less than 15%, and projects the balance for 30 days. Remaining spending is modeled as independent daily draws from the last 60 days of history, so the 80% band widens with the square root of the horizon \(d\):

where \(\hat\mu\) is the mean daily discretionary flow over the last 60 days (negative, because only outflows enter it), \(\hat\sigma\) its standard deviation, \(R_d\) the sum of recurring flows due by day \(d\), and \(X_d\) the sum of known one-off flows such as pending refunds.

Values guardrails

Members can turn on guardrails for gambling, alcohol, tobacco and vaping, adult content, and crypto and trading apps. Enforcement has two tiers, which the product never blurs: network blocks, where the card program declines a merchant category at authorization, and detection, where the engine recognizes a merchant after the charge and then pauses the card, notifies an accountability partner the member chose, or surfaces support resources such as a problem-gambling helpline. Switching a guardrail off takes effect only after a cooldown (seven days for gambling, two to three days for the others), so the guardrail is a commitment chosen in a calm moment rather than a setting undone in a hard one. Self-exclusion programs built on a similar idea are under-used and do not stop all gambling, although participants generally report benefits, including less gambling .

A worked example

Figures {{n:fig-timeline}} and {{n:fig-factors}} show the engine’s output on the synthetic member behind our website: 23 days of card transactions (September 12 to October 4, 2026, New York time; the synthetic dates run past this paper’s date), three receipts, and six health signals for “Alex,” whose Sunday nights are hard. Groceries, pharmacy, and fuel score 0 even after short nights. A $6.50 coffee the morning after a poor night scores 25, below the elevated threshold, because context counts half on small tickets. The engine finds all three cycles: a regret loop (two stressed purchases totaling $306, both returned within 45 days), a recurring window (Sunday nights: six stressed purchases across four Sundays, $566), and a binge (three purchases in six hours, $244). At 11:48 p.m. on the fourth Sunday, a $128 order from a merchant whose previous order Alex had sent back scores 83. The engine matches the order-confirmation email to the charge with confidence 1.0, proposes a $125 per-purchase limit until 8 a.m. (110% of $112.45, the largest of Alex’s nine essential purchases, which the rule in Equation 6 selects when there are ten or fewer, rounded up to $5), and schedules a return reminder three days before the stated return window closes.

The engineering harness

The engine is a set of pure functions composed by a single entry point, understand(member), which matches receipts, scores purchases, detects cycles, and plans interventions. Card events arrive by signed webhooks (verified under the Standard Webhooks scheme and deduplicated), receipts by an inbound-email route, and health deltas by an authenticated route that discards any field other than the three deltas. In auto mode, interventions execute through the card platform’s API; in suggest mode they are shown to the member first. A scheduled job reverts every temporary limit and pause. The 45 unit tests cover receipt parsing, matching, scoring, cycle detection, intervention planning, guardrails, giving calculations, and webhook verification.

From hand-set weights to learned models

The version-0 weights are priors, chosen from the literature in Section 7 and from product judgment. Once members consent to outcome measurement, we plan to learn per-member weights. Our candidate is a generalized additive logistic model of the probability that a purchase is later regretted, with regret labeled by returns of the item, by the member’s own answers in the app, and by declined Calm Mode suggestions treated as disagreement. An additive model keeps the property we care about most: each prediction decomposes into per-factor contributions that can be shown to the member. Member-level random effects let the model learn that Sunday nights are hard for one person and Friday afternoons for another; monotonicity constraints keep explanations sensible (more sleep debt never lowers a score); and thresholds will be set to bound the rate of interventions on purchases members later endorse. Race, ethnicity, gender, disability, and direct stand-ins for them, such as names or neighborhoods, will not be features. Other features, including health signals and time of day, are correlated with these characteristics, so any learned model would face the same audit as version 0 (Section 14.2). Language models will be used where language is the task, such as extracting receipts and answering members’ questions, not to score people.

Planned evaluation and ethics

What we plan to measure

We plan to register the design and the outcomes in Table {{n:tab-evaluation}} in a public registry before launch. Members who opt in to Calm Mode would be randomized, for a fixed period, between suggest mode and a holdout in which scores are computed but no interventions are shown, and analyzed by intention to treat. New intervention types would roll out in steps so that each can be evaluated against the others.

Status. This is a plan, not a registered protocol. Nothing has been registered and no registry has been chosen; no institutional review board or other independent ethics committee has reviewed the design; and the sample size, minimum detectable effects, unit of randomization, stopping rules for harm, data-retention periods, and consent language have not been set. Members assigned to the holdout would not receive interventions they might expect from Calm Mode, so the consent language would need to say so plainly before randomization.

Fairness

Because the burden of stress spending is unequal, its benefits and harms might be too. We plan to report outcomes by income band and age band and, for members who choose to share them, by self-reported gender, race and ethnicity, disability, and shift work, the last because the time-of-day factor may fall more often on people who work at night. These answers would be collected only for auditing and stored apart from every decision system. We would check that interventions do not fire more often for some groups than others at equal underlying risk, that the essentials floor protects lower-income members, whose essential purchases are smaller, and that the calm-week cashback does not pay less to members whose health signals differ because of illness, disability, or medication.

Privacy, autonomy, and paternalism

Health data is sensitive. Whether the Federal Trade Commission’s Health Breach Notification Rule or state consumer health-data laws apply to Vueni is a legal question that this paper does not settle; we treat health deltas as sensitive in storage, logs, and analytics regardless. Members can turn Calm Mode off at any time; suggest mode is the default; every intervention explains itself and reverts automatically; and the only friction that persists is friction the member chose in advance. We will not sell data, use health data for credit or advertising, or optimize for engagement.

What we will publish

We will publish the registered protocol; the outcome results, whether or not they favor the product; the distribution of interventions across groups; and every change to the model’s weights and thresholds.

Part II has described what we built, how it works, and how we plan to learn whether it helps. We invite scrutiny of all three.

Data and code for Part II. The reach targets of Section 12 are computed by wealthsim.mjs (its reach phase) in the replication files, under the planning assumptions stated there. The worked example of Section 13 is the Understanding engine’s output on a synthetic member, included in the replication files as engine-demo.json; the engine’s source code is proprietary.

References