Dynamic-Regime-Allocation allocates between the Nifty 50 and cash. Rather than forecasting prices, it uses an unsupervised Gaussian Hidden Markov Model to infer which of three latent regimes — bull, sideways, bear — the market is currently in, and sizes exposure from those probabilities.
The model is fit walk-forward: every 10 trading days it is refit on the trailing 5 years and then scores only the 10 days that follow. No parameter, scaler, or state label is ever derived from a bar the model has already been asked to trade.
| Metric | Strategy | Buy & hold |
|---|---|---|
| Total return | 217.68% | 302.95% |
| CAGR | 8.91% | 10.83% |
| Volatility (ann.) | 11.12% | 16.18% |
| Sharpe (vs cash) | 0.30 | 0.36 |
| Sortino | 0.35 | 0.46 |
| Max drawdown | -23.89% | -38.44% |
| Calmar | 0.37 | 0.28 |
| Longest underwater | 776 d | 741 d |
| Worst day | -5.93% | -12.98% |
| Time in market | 69.01% | 100% |
| Trades | 65 | — |
| Costs paid | 6.50% of capital | — |
Read this honestly. The strategy does not beat the index. It gives up about 2 points of CAGR and lands slightly behind on Sharpe. What it buys is a materially smaller hole: max drawdown falls by a third, the worst single day is less than half as bad, and Calmar — return per unit of worst-case pain — is the one headline measure where it wins. Whether that trade is worth making is a question about the holder, not about the model.
The COVID crash is the model's best moment and it is a genuine one. It exited on
2020-02-05 at 12,089 — three sessions after P(Bear) crossed the threshold,
and within 2% of the all-time high — and stayed in cash through the bottom.
| Feb–Apr 2020 drawdown | |
|---|---|
| Buy & hold | -38.4% |
| Strategy | -5.3% |
Three scale-free inputs, so a window from 2013 (Nifty ~6,000) and one from 2026 (Nifty ~24,000) live on the same axes:
- Volatility — 20-day annualised rolling stdev of returns.
- Trend (Z-score) — price distance from its 50-day mean, in sigmas.
- Momentum — trailing 14-day return.
Refit every 10 days on a trailing 1,250-day window, best-of-3 EM restarts by in-sample log-likelihood. Latent state labels are arbitrary and permute between fits, so states are re-identified each window by their mean trend feature and ordered bear → sideways → bull.
A 10-day moving average is applied to the regime probabilities, so a signal must persist for about two weeks before capital moves. This is the main churn control: it holds the strategy to 65 trades across 13 years.
- Friction — 0.1% per unit of exposure traded (brokerage + STT + slippage).
- Cash yield — un-invested capital earns 6% annualised (liquid/overnight funds).
- Financing — exposure above 1x is charged 6% + 2% spread.
| Regime | Condition | Exposure |
|---|---|---|
| Strong bull | P(Bull) > 80% | 1.5x (off by default — see below) |
| Bull | P(Bull) > 60% | 1.0x |
| Sideways | P(Side) > 60% and Z > -0.5 | 1.0x |
| Bear | P(Bear) > 60% | 0x |
The rule is deliberately sticky: once invested, only a bear reading returns it to cash. Exiting on "not bullish enough" was what produced most of the churn. The 80% / 75% pair forms a hysteresis band so the top tier cannot flip on and off around a single threshold.
The aggressive tier ships disabled (leverage_max = 1.0). It is implemented
and correct — enable it with --leverage-max 1.5 — but on this data it makes
every risk-adjusted measure worse once financing and the extra turnover are
charged for:
| 1.0x (default) | 1.5x tier enabled | |
|---|---|---|
| CAGR | 8.91% | 7.94% |
| Sharpe | 0.30 | 0.20 |
| Max drawdown | -23.89% | -28.12% |
| Calmar | 0.37 | 0.28 |
| Trades | 65 | 159 |
| Costs paid | 6.50% | 11.25% |
git clone https://github.com/SahilMotyar/Dynamic-Regime-Allocation.git
cd Dynamic-Regime-Allocation
pip install -r requirements.txt
python hmm.pyEvery assumption is a flag; python hmm.py --help lists them all.
# a different index, and no chart window
python hmm.py --ticker ^GSPC --start 2005-01-01 --no-plot
# stress the cost assumptions
python hmm.py --tx-cost 0.003 --cash-rate 0.04
# enable the aggressive tier, save the outputs
python hmm.py --leverage-max 1.5 --save-plot docs/levered.png --save-csv outputs/levered.csvPrices are cached under data_cache/ and reused for the rest of the day; pass
--refresh to force a re-download.
pip install -r requirements-dev.txt
pytest49 tests, no network access required — the model tests run against a synthetic two-regime price path.
The run ends with a trade card for the next session. It reads the target exposure straight out of the backtest's final row, so the recommendation is by construction the same rule that produced the table above.
==========================================================
LIVE SIGNAL - 2026-08-13
==========================================================
Last close 24,395.85
Trend (Z-score) 0.84 sigma
Volatility (ann.) 10.3%
----------------------------------------------------------
Regime beliefs (smoothed):
Bear 0.4%
Sideways 0.1%
Bull 99.5%
----------------------------------------------------------
Currently held 1.00x
Target exposure 1.00x
Regime INVESTED - standard, 1.00x
Action HOLD
==========================================================
- The backtest starts in 2013, not 2007. The first 1,250 trading days are
consumed by the initial training window. Earlier versions padded that warm-up
with a uniform 1/3 prior, which kept the strategy in cash — earning 6% — right
through 2008. That produced a flat line across the crash that looked like
regime detection but was just an untrained model sitting in cash. The warm-up
is now
NaNand those rows are dropped. - One asset, one market. Results on Nifty over one 13-year sample; the thresholds have not been validated out-of-sample on other indices.
- Thresholds are hand-set, not fit. They have not been walk-forward optimised, which is a virtue (nothing is fit to the test period) and a limitation (nothing says they are near-optimal).
- No slippage model on gaps. Exits are filled at the close of the signal day; a genuine crash open would fill worse.
- Survivorship and index reconstitution in the underlying Nifty series are taken as given from the data provider.
regime_allocation/
config.py dataclasses holding every tunable parameter
data.py download, on-disk cache, feature engineering
model.py walk-forward HMM, state ordering, belief smoothing
strategy.py the allocation rule (shared by backtest and live signal)
backtest.py position lagging and net-of-cost PnL accounting
metrics.py risk/return statistics and the comparison table
reporting.py trade card and charts
cli.py argument parsing and orchestration
hmm.py entry point: python hmm.py
tests/ 49 tests, no network required
Educational and research use only. This is not financial advice. Algorithmic trading carries significant risk, and past performance does not indicate future results.