We audited our own model the way a skeptic would: refit quarterly using only data available at the time, signals applied the next day, costs included. Across all 20 markets. These are the results — including where the model loses.
The honest test. Most published backtests train a model on the full history and then "predict" that same history — the model has already seen the answers. Here the HMM is refit every quarter (~53 refits per market) using only past data, then classifies the next quarter it has never seen. The day-T signal decides the day-T+1 position, with 0.10% cost per switch. This is called walk-forward validation.
—
Bear-day vs bull-day volatility
Next-day realized vol on days labelled bear vs days labelled bull, out of sample. This is the model's real skill.
—
Markets with smaller max drawdown
Stepping to cash in bear regimes cut the worst peak-to-trough loss in most markets.
—
Markets beating buy & hold on Sharpe
As a return-timing strategy it mostly loses. We publish this anyway — it tells you what the model is for.
—
Label agreement, honest vs full-history
How often the out-of-sample regime label matches the dashboard's full-history label — a stability check.
What this means for how you use the dashboard. Treat the HMM as a sea-state gauge, not a compass. A bear label reliably means rougher waters ahead — higher volatility, fatter left tail — so it's a signal to size down, hedge, or tighten risk. It is not a reliable signal that prices will fall: in some markets (Nifty included) bear-labelled days actually saw higher average next-day returns, because panic days cluster with rebound days. Vol is forecastable; direction largely isn't.
All 20 markets — walk-forward strategy vs buy & hold
Strategy = long the index unless the regime is bear, then cash at the local risk-free rate, net of costs. Green = strategy beat buy & hold on that metric. Sorted by Sharpe improvement.
Market
Strategy CAGR
B&H CAGR
Strategy Sharpe
B&H Sharpe
Strategy max DD
B&H max DD
Bear/bull vol
Market detail
——
Growth of 100 · out-of-sample window · strategy shown net of 0.10% per switch.
Does today's label predict tomorrow?
Today's regime
% of days
Next-day vol (ann.)
Next-day return
Days falling >1%
Strategy behaviour
% time invested
—
rest in local cash
Switches
—
index ↔ cash, whole window
Avg regime length
—
trading days
Label agreement
—
vs full-history labels
Method, in full. 3-state Gaussian HMM per market on daily log returns, 10-day realized vol, FX change and (where available) implied vol — the same features the live dashboard uses. Initial training window ~3 years, refit every 63 trading days on an expanding window; features are standardised using only the training window. States are labelled bear/neutral/bull by a volatility-dominant score, and probabilities are filtered (causal), never smoothed. See Methodology for the model itself. Results regenerate when we rerun the audit; last run —. Not investment advice — see Disclaimer and Terms.