Does raising retirement contributions increase net wealth accumulation and improve lifetime welfare? The answer depends on how individuals finance the additional contributions: by cutting spending, reducing non-retirement savings, or increasing debt. We use newly linked deposit, credit, and pension data from a large UK financial institution to study a gradual increase in minimum retirement contributions from 2% to 8% of salary. On average, only one-third of the increase is funded through reduced spending (especially in discretionary categories like restaurants and leisure), with the rest of the decrease in take-home pay financed through higher borrowing and lower non-retirement saving. Low-liquidity individuals primarily cut spending, while those with high-liquidity shift existing savings across accounts with minimal impact on spending over two years. We show that these behavioral responses are crucial for welfare analysis in a sufficient statistics framework: well-targeted policies raise contributions among those who (i) exhibit less crowd out and (ii) are more prone to undersaving (e.g., due to present bias). We identify these groups in a lifecycle model by targeting the observed spending responses (to identify low crowd-out individuals) and the persistence of simultaneous retirement saving and credit card borrowing (to identify present bias). Using the estimated model, we find that standard financial incentives for retirement saving are poorly targeted, as they are disproportionately taken up by high-liquidity, time-consistent individuals who exhibit high crowd-out. Conversely, asset tests on tax incentives help screen out such individuals, and retirement income floors and annuities provide commitment for present-biased households.

We develop and implement a novel method to study personal financial advice from Large Language Models (LLMs). Studying this advice is challenging because it depends on the model used (i.e., supply), the questions individuals ask (i.e., demand), and their evolving circumstances. We address these challenges by surveying a representative sample of adults and asking them to write prompts seeking spending and investing advice from an LLM. We then simulate the lifetime paths that result from following this advice under realistic asset and labor market conditions. Applying our method to GPT-5.2 and Gemini 3.0 Flash, we document three facts about AI-generated financial advice. First, following LLM advice would move most survey respondents closer to the prescriptions of life cycle theory relative to their current behavior, including broader participation in diversified equity funds, equity shares that decline with age, and sizeable saving buffers. Second, replacing individual-written prompts with academic prompts moves LLM advice even closer to life cycle theory, with better consumption smoothing and less reliance on simple heuristics. Third, LLM advice varies systematically with individual characteristics, such as gender and financial literacy. These differences accumulate over the life cycle into wealth differences at retirement of 4-5% between groups and reflect both demand (i.e., systematic variation in the prompts written by different individuals) and supply (i.e., differences in advice for a given prompt). These facts highlight the potential of generative AI to improve financial decision-making, but suggest that its impact is likely heterogeneous across households and depends on how the technology is used.

Household subjective expectations are typically studied one domain at a time, but their effects on behavior depend on their joint distribution: for example, optimism about stock returns may raise equity shares, but if optimism extends to all assets, portfolio shares may remain unchanged. We measure this joint distribution using a new two-wave panel survey that elicits expectations about asset returns, labor market outcomes, inflation, and life expectancy for a representative sample of U.S. adults. We find that beliefs co-move along a small number of common factors, sort individuals into psychologically interpretable types, and display substantial within-person persistence once measurement error is corrected for. We incorporate this belief structure into a lifecycle model of consumption, saving, portfolio choice, and housing to quantify the behavioral and welfare implications of correlated expectations. Relative to a benchmark with equally dispersed but uncorrelated beliefs, the observed correlation structure substantially attenuates the pass-through from any single expectation to behavior and reduces the average welfare cost of belief heterogeneity by nearly one-third. Studying expectations one domain at a time thus overstates the behavioral and welfare consequences of belief heterogeneity.

Evaluating reforms requires predicting how individuals would behave under counterfactual policies and mapping those behavioral responses into welfare-relevant criteria. We use survey responses to hypothetical scenarios linked with administrative 401(k) data to address both in the context of designing employer matching formulas. We find that (i) these survey responses can accurately predict saving responses in administrative data, (ii) saving is inelastic to the match rate, and (iii) non-elective contributions do not crowd out employee saving. These patterns imply that (iv) plans combining lower match rates with non-elective contributions generate higher savings and more equitable match distributions, and (v) many existing plans—including safe-harbor formulas—are dominated along both dimensions. These conclusions hold for any objective function that values either higher saving or lower inequality, or both.

U.S. employers and the federal government devote the equivalent of 1.5% of GDP annually toward promoting defined contribution (DC) retirement savings. Using a new employer-employee linked dataset covering millions of Americans, we show that tax and employer matching incentives disproportionately benefit White and Asian workers compared to their similar-income Hispanic, Black, and American Indian or Alaska Native coworkers. Similarly, these incentives disproportionately benefit those with richer parents compared to those from lower-income families. Breaking the link between contribution choices and saving subsidies through revenue-neutral reforms could close up to one-third of the DC wealth gaps by race and parental income.

We study the role of risk preferences and frictions in portfolio choice using variation in 401(k) default investment options. Patterns of active choice in response to different default funds imply that, absent participation frictions, 94% of investors prefer holding stocks, with an equity share of retirement wealth declining with age—patterns markedly different from their observed allocations. We use this quasi-experiment to estimate a lifecycle model and find relative risk aversion of 2.5, EIS of 0.25, and a $160 portfolio adjustment cost. Our results suggest low stock-market participation is due to participation frictions rather than non-standard preferences such as loss-aversion.

Using data from over 100 U.S. retirement plans and a representative U.K. panel, I document the impact of auto-enrollment on retirement savings at different horizons. I replicate the impact of auto-enrollment on participation and contributions in the short–run (at 12 months), but I show that these gains are attenuated over the medium-run (at 36 months). At this longer horizon, the average savings increases are modest, though auto-enrollment significantly lowers inequality in savings. To assess auto-enrollment’s lifetime impact, I estimate a lifecycle consumption-savings model that can fit the observed patterns with a switching cost of approximately $250, smaller than previous estimates.

We study how couples allocate retirement-saving contributions across each spouse’s account. In a new dataset covering over a million U.S. individuals, we find retirement contributions are not allocated to the account with the highest employer match rate. This lack of coordination—which goes against the assumptions of most models of household decision-making—is common, costly, persistent over time, and cannot be explained by inertia, auto-enrollment, or simple heuristics. Complementing the administrative evidence with an online survey, we find that inefficient allocations reflect both financial mistakes as well as deliberate choices—especially when trust and commitment inside the households are weak.

We investigate whether the ‘one-child policy’ has contributed to the rise in China’s household saving rate and human capital in recent decades. In a life-cycle model with intergenerational transfers and human capital accumulation, fertility restrictions lower expected old-age support coming from children—inducing parents to raise saving and education investment in their offspring. Quantitatively, the policy can account for at least 30% of the rise in aggregate saving. Using the birth of twins under the policy as an empirical out-of-sample check to the theory, we find that quantitative estimates on saving and education decisions line up well with micro-data.