R&D · Research noteComplete

Regime-Adaptive Equity Model

Regime-adaptive equity strategy research on Indian markets (NIFTY 200), as much about disciplined negative-result reporting as the final strategy.

QuantFinanceResearchHMM
ScopeResearch
DomainQuant
StatusComplete

Key metrics

13.5%

Strategy CAGR

Net of friction, 2024–2026

7.7%

Buy-and-hold CAGR

NIFTY baseline

0.83

Sharpe

vs 0.65 buy-and-hold

64%

Win rate

vs 52% baseline

9.4 pp

2008 drawdown cut

HMM regime sizing ablation

Research question

Most retail-adjacent quant research falls for its own backtest: a dashboard shows gross returns, looks great, and the friction, regime risk, and realistic execution constraints get added later, if at all. Building a strategy that survives contact with real costs and real regime shifts requires being willing to kill your own architecture when the evidence says so.

Method

Phase 1 built an XGBoost baseline and found that daily-return prediction is mostly noise, the real signal is monthly. Phase 2 built a custom multimodal transformer (regime cross-attention, mixture-of-experts, ranking loss) to exploit that signal directly; a 60-line diagnostic showed it performing worse than plain 21-day momentum with zero ML.

Phase 3 pivoted: production became a 21-day momentum ranking with 3-state HMM regime-based position sizing, sector caps, per-stock stops, and a portfolio drawdown killswitch, benchmarked with 0.22%/rebalance friction. Not a high-frequency system, rebalances every 21 trading days.

Architecture

  1. 01

    Phase 1, XGBoost baseline; daily-return prediction mostly noise, monthly signal holds

  2. 02

    Phase 2, Custom multimodal transformer (RAMT); diagnostic showed it underperformed plain 21-day momentum

  3. 03

    Phase 3, Production: 21-day momentum ranking + 3-state HMM regime sizing (100%/50%/20%)

  4. 04

    Risk controls, sector caps, per-stock stops, portfolio drawdown killswitch, 0.22%/rebalance friction

  5. 05

    Validation, 4-window historical ablation (2008, 2010, 2013, 2024)

Findings

  • Production strategy backtest (2024–2026, net of friction): 13.5% CAGR vs 7.7% NIFTY buy-and-hold, Sharpe 0.83 vs 0.65, 64% win rate vs 52%
  • Diagnostic script exposed the custom transformer underperforming a zero-ML momentum sort on every metric, published, then rebuilt around the finding
  • Regime-based position sizing cut historical 2008-style drawdown by 9.4 percentage points in ablation testing, at the documented cost of capping upside in bull markets

Method details

Universe
NIFTY 200 equities
Strategy
21-day momentum ranking + 3-state HMM regime sizing (100%/50%/20% allocation)
Also tested
Custom transformer (RAMT), Chronos-T5 + LoRA foundation-model fine-tune
Backtest rigor
Realistic friction, per-stock stops, portfolio drawdown killswitch, 4-window historical ablation (2008, 2010, 2013, 2024)

Stack

AI / ML
PyTorchChronos-T5LightGBMXGBoostHMM
Data
yfinancepyarrow
Delivery
StreamlitIEEE-format writeup

References

  • NIFTY 200 equities universe (yfinance)
  • Chronos-T5 + LoRA foundation-model fine-tune (ablated)
  • IEEE-format research writeup + Streamlit dashboard

Applying this method in production?

Talk research → product