HMM + RL regime allocation paper (SPY/TLT/GLD) — hype post, weak paper — tweet
HMM + RL regime allocation paper (SPY/TLT/GLD) — hype post, weak paper — tweet
· @0xTrackmind
Tweet
Source claim (post, 2026-09-22): a quant paper combining a Hidden Markov Model with reinforcement learning 'beat SPY on risk-adjusted returns, with shallower drawdowns across 2004-2025'. The paper is arXiv 2605.27848 (Verma, Putri, Lesupi; independent researchers; dated Nov 2025, posted 2026-05-27; not peer-reviewed; no code). Desk read: its own tables undercut it — a plain monthly equal-weight SPY/TLT/GLD portfolio has the best full-sample Sharpe (0.80 vs 0.52 and 0.58 for the HMM rotations, 0.46 SPY) and ties the RL policy out of sample (0.83 each) at 11.0% vs 17.3% volatility; SPY had the highest full-sample return. The 'RL' policy is three rows (SPY in calm and middle states, gold in the crisis state), and because the action cannot change the next regime, policy iteration reduces to picking the highest expected return per state, yet the reported policy differs from the paper's own top-return table. Zero transaction costs while a column is labelled net; whether the backtest used causal filtered regimes or full-sample Viterbi/smoothed ones is unstated. Two references could not be found (a 2023 Journal of Finance paper by Baele, Bekaert and Inghelbrecht; Zaremba and 'Karel T.' Idzorek 2023). Not a runnable published rule; nothing to import. One takeaway for challenger replays: carry an equal-weight SPY/TLT/GLD benchmark. Account: 'ai & onchain research', 3.6k followers; the po...
Takeaway
Source claim (post, 2026-09-22): a quant paper combining a Hidden Markov Model with reinforcement learning 'beat SPY on risk-adjusted returns, with shallower drawdowns across 2004-2025'. The paper is arXiv 2605.27848 (Verma, Putri, Lesupi; independent researchers; dated Nov 2025, posted 2026-05-27; not peer-reviewed; no code). Desk read: its own tables undercut it — a plain monthly equal-weight SPY/TLT/GLD portfolio has the best full-sample Sharpe (0.80 vs 0.52 and 0.58 for the HMM rotations, 0.46 SPY) and ties the RL policy out of sample (0.83 each) at 11.0% vs 17.3% volatility; SPY had the highest full-sample return. The 'RL' policy is three rows (SPY in calm and middle states, gold in the crisis state), and because the action cannot change the next regime, policy iteration reduces to picking the highest expected return per state, yet the reported policy differs from the paper's own top-return table. Zero transaction costs while a column is labelled net; whether the backtest used causal filtered regimes or full-sample Viterbi/smoothed ones is unstated. Two references could not be found (a 2023 Journal of Finance paper by Baele, Bekaert and Inghelbrecht; Zaremba and 'Karel T.' Idzorek 2023). Not a runnable published rule; nothing to import. One takeaway for challenger replays: carry an equal-weight SPY/TLT/GLD benchmark. Account: 'ai & onchain research', 3.6k followers; the post is a paper-hype template ('bookmark it before this thread gets buried'), 231 views and 0 likes at capture.
Related
3 eventsSource links
Tweets, videos, filings, articles, and source pages tied to this read.