Lookahead Bias from Smoothed Regime Estimates
This failure mode arises when an HMM backtest labels each period’s regime using full-sample smoothed probabilities — which condition on the entire observation sequence, including future data — and then “trades” those labels as if they had been known in real time. Because smoothed estimates use information from after the trade date, the strategy appears prescient when it is in fact peeking. Honest evaluation requires filtered or online inference, where the regime for day t uses only data up to day t; Bulla et al. (2011) and Shu/Yu/Mulvey (2024) both explicitly use rolling-window online inference for this reason. It appears in this vault as a key reason many HMM regime “profitability” claims do not survive proper out-of-sample testing.
Connections
- Hidden Markov Model Regime Detection — suffers_overfitting_risk, source: https://mpra.ub.uni-muenchen.de/21154/1/MPRA_paper_21154.pdf
- Out-of-Sample Backtesting — contradicts, source: https://arxiv.org/html/2402.05272v2
- Zakamulin 2016 — supports, source: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2743119
- Regime Classification — contradicts, source: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2743119