Papers & software

We embed a two-asset HANK household problem in a stock-flow-consistent agent-based model of the Dutch economy. Households optimise consumption and portfolio decisions under subjective market beliefs, drawing on Adam and Marcet's approach. An amortised neural-network solver makes these decisions computationally feasible across simulations. Calibration uses national accounts and administrative microdata from Statistics Netherlands, retaining heterogeneity in income, liquid wealth and illiquid wealth. Rolling out-of-sample forecasts compare the model with heuristic household rules and time-series benchmarks under four expectation specifications. At the twelve-quarter horizon, root mean squared forecast errors of aggregate household consumption are 40.5–54.2% lower than matched univariate benchmarks, and those of investment 14.5–15.5% lower; with 28 overlapping forecast origins, only the investment gains are statistically significant. The consumption gain is concentrated in the 2013–14 consumption trough, which the model captures while staying within 10.2% of the benchmark afterwards. Comparisons with heuristic household rules point to dynamic household behaviour and state-dependent portfolio choice as sources of the gain. The paper examines when optimising households improve macroeconomic forecasts and which decision mechanisms account for those gains.

Can estimation distinguish a change in cost pass-through from compensation for a wrong wage-setting rule? I study this in a Dutch agent-based economy built on CANVAS, with a common markup pricing rule and two wage rules, price indexation and error correction, estimated by neural simulation-based inference. In a known-truth experiment with 100 independent panels per cell, fitting the indexation model to error-correction data recovers a break in the cost pass-through coefficient almost without bias, but overstates its level and overstates the twelve-quarter consumer-price response to a wage shock; fit and forecast warnings flag every such panel. I apply the framework to the 2021–23 Dutch inflation episode.

BeforeIT.jl is an open-source Julia framework for macroeconomic agent-based modelling, implementing the model of Poledna et al. (2023). Its modular structure supports economic forecasting, counterfactual simulations and model extensions. The paper presents the framework, worked examples and performance benchmarks. I implemented the CANVAS and growth-rate model variants and the multicountry forecasting pipeline with its validation tools.

How much of the welfare loss from the Dutch housing crisis reflects too few homes, and how much reflects households being poorly matched to the existing stock? I am building a new macroeconomic ABM that combines elements of BeforeIT.jl and Wiese et al. (2024), with spatial housing markets, job search and job-to-job mobility in the tradition of Postel-Vinay and Robin, demography, firm entry and exit, and endogenous growth. Housing mechanisms include rent and mortgage lock-in, underwriting constraints and construction delays. The project is designed to separate shortage from misallocation over 2015–2035 and compare policies that expand or reallocate the housing stock, including the mortgage interest deduction, rent regulation, nitrogen limits on construction and grid congestion. Model integration, calibration and welfare comparisons are ongoing.

When does richer microdata on the households and firms a structural model starts from improve its counterfactuals? A framework spanning agent-based, heterogeneous-agent and panel models.