Ask why a strategy stopped working and the honest answer is usually not "the strategy broke" — it's "the market changed regime and nobody was watching."
Two strategies, opposite assumptions
Trend-following assumes price movement persists: a breakout tends to keep breaking out. Mean-reversion assumes the opposite: an extreme move tends to snap back. Both assumptions are true — just in different markets. A trending, low-noise market rewards trend-following and punishes mean-reversion. A choppy, range-bound market does the reverse. Running either strategy unconditionally across all market conditions guarantees it eventually fights the tape.
What a regime actually is
A useful regime classification isn't just "bull or bear." The dimensions that matter for strategy selection are:
- Directionality — is price trending, or oscillating around a level?
- Volatility — is the range expanding or compressed?
- Persistence — how long has the current condition held, and how often does it flip?
A strategy validated only in one regime combination (say, trending + low-volatility) will falsify itself the moment the market spends a month ranging — not because the code is wrong, but because the precondition for the edge no longer holds.
The mistake regime-detection tries to fix
The common failure mode isn't picking the wrong strategy — it's picking the right strategy for the wrong period, without a mechanism to notice the mismatch. A dip-buying strategy that works beautifully in a RANGING regime can bleed steadily in a TRENDING_BEAR regime, because "buy the dip" assumes the dip resolves upward, and in a real downtrend it usually doesn't.
Building regime awareness into a strategy, not around it
The fix isn't a single "master indicator" that predicts regime perfectly — nothing does that reliably. It's:
- Classify the current regime from evidence available at decision time — price structure, volatility, and (where it exists) broader market context — never with hindsight.
- Gate strategies to the regimes they were actually validated in. A strategy backtested only in RANGING markets should not be live in TRENDING_BULL just because it's "on."
- Re-validate periodically. Regimes shift, and a strategy's fit to a regime can decay even if the regime label hasn't changed.
- Treat regime tagging as a proxy, not ground truth, and say so — a deterministic OHLC-based regime classifier used for backtesting will disagree with a live, context-aware classifier sometimes, and pretending otherwise overstates confidence in the backtest.
This is the same principle behind why Atradian tags every backtest with a per-regime breakdown instead of a single blended number, and why the live bot gates which strategies are even eligible to trade based on the current classified regime — a strategy's edge is a claim about a specific market condition, not a universal truth.