Imagine testing a driving strategy only on sunny highway days. Never rain. Never ice. Never city traffic. The strategy looks brilliant.
Now deploy it in February. Black ice. Tram stops. Kids running into the road.
That's regime bias in trading.
What Regime Bias Is
A market regime is a period where certain patterns dominate. Examples:
- Bull market: sustained upward trend, low volatility, "buy the dip" works
- Bear market: sustained downtrend, high volatility, "buy the dip" is a graveyard
- Sideways: no trend, mean reversion works, momentum fails
- Panic: correlations spike to 1, everything crashes together, diversification fails
- Mania: everything goes up regardless of quality, "obviously overvalued" wins
A strategy designed in one regime often fails in another. If your backtest covers only one regime, you can't know.
The Crypto Example
Bitcoin from 2016 to 2021 was basically a bull market with one interruption (2018). 85 percent of all "Bitcoin trading strategies" ever published were tested in that window.
Which means they were tested on a single regime: long bull with a brief bear.
Apply those same strategies to 2022 (full bear) or 2015 (long sideways). Most collapse.
This is why every BotLab experiment on BearBullRadar uses minimum 6 years of data, covering the 2018 bear, the 2020 COVID crash, the 2021 bull, and the 2022 bear. Four distinct regimes minimum.
Strategies that survive all four have a chance at being real.
My Sandra Example
Her strategy was tested from 2015 to 2026. That's 11 years.
Regimes it covered:
- 2015-2018: long bull (trading from $2k to $10k not including crypto)
- 2018: brief bear (actually not as bad for big-cap stocks)
- 2020: COVID crash and recovery
- 2022: growth stock bear
- 2023-2025: AI-driven bull
That's 5 regimes. Not bad.
But — and this is the nuance — US equities from 2015 to 2026 have been mostly bullish. The 22 year period 2000-2022 had two full bears (2000-2002 dotcom, 2007-2009 GFC) each lasting 2-3 years.
Sandra's test didn't include a prolonged bear. A strategy that deploys capital on every 10-percent drop in a 3-year bear would deploy 5 tranches and still watch prices fall another 40 percent. Capital destruction.
Her strategy didn't face that test. The backtest doesn't show it. But it's a risk in the future that can't be ruled out.
The Momentum Regime Problem
Momentum, the strategy I actually recommend, also has regime problems.
Documented weak periods for momentum:
- 2008-2009: crisis crashes everything, including winners
- 2016: value rally after Trump election, momentum underperformed for 18 months
- March 2020: COVID panic crashed momentum stocks hardest
- 2021 meme stock squeeze: GameStop-style reversals hit shorts (if applicable)
If you deploy momentum next year and it underperforms SPY by 10 percent, don't panic. That happens. The long-term edge holds through regime shifts, but not every year.
How to Defend Against Regime Bias
1. Demand 10+ years of data. Minimum. Preferably 20+.
2. Identify the regimes in your test window. If you can't name 3+ distinct regimes, your test is too short.
3. Check behavior per regime. Does the strategy work in bull only? Bear only? Sideways? If the answer is "only in one," caveat emptor.
4. Stress-test on the missing regime. My Sandra test didn't include 2000-2002 or 2007-2009. If I wanted to stress-test, I could fire up historical data and simulate.
5. Understand WHY a strategy works. Momentum works because of behavioral biases (slow reaction to news, herding into winners). Those biases exist in all regimes, but express more strongly in some. Knowing the mechanism tells you when to expect underperformance.
What You Should Expect
Any real strategy has bad years. Not every year is alpha.
- SPY has had years down 30+ percent
- Value strategies had a decade of underperformance (2009-2019)
- Momentum had the 2008-2009 disaster
- Buy-and-hold on good stocks still had 50+ percent drawdowns
If a strategy promises "every year beats the market," it's regime-biased or overfit. No real strategy does that.
The bar: over 10+ years, does it beat the benchmark after costs?
If yes, you can tolerate bad years. If no, you should use the benchmark.
The Practical Takeaway
Check the regime composition of any backtest you're shown.
"This strategy made 40 percent in 2022" — cool, 2022 was one regime. How about 2020? 2018? 2015?
"Momentum has beaten SPY since 1927" — now we're talking. 98 years covers every possible regime.
"My strategy has a 5-year track record" — marginal. One regime, maybe two.
The shorter the track record, the more regime-biased the claim.
-> Previous: P-Hacking -> Next: Recency Bias -> Back to pillar
Sources
- Ang, Goetzmann, Schaefer — Evaluation of Active Management — regime-aware factor evaluation
- Asness, Frazzini, Pedersen — Buffett's Alpha — how Buffett's style performs across regimes
Your Dominic, who insists on 10-year backtests so you don't learn the hard way.
Disclaimer: Not financial advice. Past performance does not guarantee future results.
