Mitigating Sequence-of-Returns Risk with Regime-Switching Asset Allocation: A Comparison of Econometric and Deep Learning Models in Portfolio Decumulation

Abstract Retirement portfolios under regular withdrawals face sequence-of-returns risk: early losses can exhaust capital even when average returns are adequate. We examine whether regime-switching dynamic asset allocation mitigates this risk for a 60/40 stock-bond portfolio, and whether deep learning identifies regimes better than classical econometrics. Five models (a Markov-switching model, a multivariate and a univariate Gaussian hidden Markov model, an LSTM, and an encoder-only Transformer) are compared under identical data, tuning, and walk-forward validation. Their signals, net of transaction costs, feed 10{,}000 stationary-bootstrap retirement paths in three withdrawal scenarios. At 4.0% the MSM, the univariate HMM, and the LSTM cut the depletion rate from 3.30% to 0.57%, 0.60%, and 0.60% with disjoint Wilson intervals; at 6.0% only the LSTM reduces it (17.01% against 33.04%), the univariate filters raising it to about 49% as their return sacrifice becomes the dominant ruin cause, and the multivariate HMM raising it in every scenario. Only the econometric models turn defensive in all five crisis windows; neither network identifies regimes more accurately, and the LSTM holds the highest median terminal capital by staying invested rather than by timing regimes. Classification quality and search diagnostics do not predict decumulation value. Regime-switching allocation in decumulation is a state-contingent tail hedge: model choice determines the protected risk profile, and above a critical withdrawal rate the premium exceeds the payout.