Regime Change and Why Strategies Decay
Strategies stop working, and they stop for reasons that are worth telling apart: the market’s statistical behaviour changed, the edge was crowded out by other participants, the structural feature the strategy relied on disappeared, or the edge was never there and the good period was noise. Only the last one is a research failure, but from inside a drawdown all four look identical — which is why the criteria for distinguishing them have to be written down before you need them.
Educational material, not trading advice. Algorithmic crypto trading is high-risk and most retail algo traders lose money.
What “regime” means
A regime is a stretch over which the market’s statistical properties are roughly stable. Volatility level, the strength of autocorrelation in returns, the correlation between assets, the ratio of trending to choppy behaviour. Every strategy is an implicit bet on some of those properties: trend-following needs returns to be positively autocorrelated over its horizon, mean-reversion needs them negatively autocorrelated, and both assume enough volatility to cover costs. See momentum vs. mean-reversion for how directly each family depends on its regime.
The important consequence is that a strategy failing in an unfavourable regime is behaving correctly. A trend follower in a choppy market is supposed to lose slowly — that’s the cost of the option it holds. Interpreting that as decay leads to abandoning something functional; interpreting genuine decay as regime drag leads to funding something dead. Both errors are expensive and they pull in opposite directions.
Four reasons a strategy stops working
Distinguishing these matters because the appropriate response differs.
Regime shift. The property the strategy exploits is temporarily absent. Diagnostic: the strategy’s performance conditioned on a measurable regime variable — realized volatility, dispersion across assets, autocorrelation of returns — matches its historical performance in that same condition. If the current losses look like historical losses under similar conditions, this is the likeliest explanation.
Crowding. Enough capital pursues the same effect that it stops being available. Diagnostic: the effect’s magnitude shrinks gradually rather than fluctuating, and it shrinks in the direction of faster reversion — the pattern gets arbitraged closer to the moment it appears. Crowding tends to be monotonic rather than cyclical, which is what distinguishes it from a regime shift.
Structural change. The mechanism the strategy relied on no longer exists: a fee structure changed, a venue’s behaviour changed, a market’s participant mix changed, an instrument’s rules changed. Diagnostic: the break is abrupt and coincides with a known event. Structural breaks are the cleanest to diagnose and the least likely to reverse.
It was never there. The backtest was noise, a leak, or the maximum of a large search. Diagnostic: the live results are consistent with zero edge from the start — and, crucially, the strength of the original evidence was weaker than it appeared. If the strategy was the best of hundreds of configurations, this explanation deserves priority over all the others, per multiple testing and strategy selection.
Why crypto is harder than most markets
The market itself has changed composition repeatedly, so the assumption that old data describes the current regime is weaker here than almost anywhere. Participant mix, dominant venue types, available instruments, typical liquidity, and the presence of professional arbitrage have all shifted substantially over the asset class’s short history.
Two research consequences follow. First, a long backtest may span several genuinely different markets, and an aggregate result across them is an average of regimes that no longer exist. Second, the sample of current-regime data is always small, which is exactly the condition under which estimates are noisy and multiple testing bites hardest.
This is the argument for rolling rather than anchored training windows: if old data describes a market that no longer exists, including it is not extra information but noise with a systematic bias. Running both schemes and comparing is more informative than choosing — a result that survives only with recent data is telling you something real about its regime dependence, per how walk-forward validation works.
Telling decay from a drawdown
The only workable method is to define, in advance, what would change your mind. After the fact, every drawdown feels like decay, and every recovery feels like vindication.
Criteria worth setting before deployment:
- A drawdown depth and duration threshold beyond anything in the validated out-of-sample history. Exceeding the backtest’s worst episode is weak evidence on its own — extremes don’t generalize, per how maximum drawdown is calculated — but a sustained excess is a genuine signal.
- A regime-conditioned expectation. Record what the strategy historically did under conditions like the present ones. Underperforming that conditional expectation is much stronger evidence than underperforming the unconditional average.
- Behavioural, not just P&L, checks. Has trade frequency changed? Has the average holding period changed? Has realized slippage widened against the model? A strategy whose behaviour has shifted has had its inputs change, which is diagnostic in a way P&L alone is not — see modelling transaction costs in a backtest.
- A pre-committed halt. A level at which the system stops regardless of your interpretation, decided while calm and enforced automatically rather than by judgement, per what sits between a backtest and a live bot.
The reason for pre-commitment is not discipline for its own sake: it’s that judgement made during a drawdown is subject to whichever bias is stronger at that moment, and both directions are available. Writing the criteria down converts a question you cannot answer objectively into one you already answered.
What doesn’t help
Re-optimizing after a bad period is the most common response and the most likely to make things worse. Refitting on data that includes the drawdown produces parameters that would have avoided it — which is fitting the recent past, by definition. If re-optimization is part of the plan, it has to be a scheduled, mechanical part of the strategy that was validated as such, not a reaction.
Adding a filter that would have sat out the bad period is the same error with a different shape: it improves the backtest and means nothing, because the filter was chosen knowing what it needed to avoid. Waiting for the regime to return is sometimes correct and unfalsifiable in the moment, which is what makes it dangerous — defensible only if you specified beforehand which measurable condition constitutes “the regime returning”.
The takeaway
Decay is the normal end state of a strategy, not an anomaly — edges are competitive and finite, and the reasonable expectation is that anything you find has a limited life. That reframing does useful work: it makes monitoring a permanent part of running a strategy rather than an occasional worry, it makes pre-committed halt criteria obviously necessary, and it lowers the emotional cost of stopping something that has stopped working.
What it does not license is the comfortable inversion — attributing every failure to a regime change. The single most likely reason a strategy that worked in a backtest does not work live remains that it never worked at all, which is why what backtesting is, and why naive backtests lie is where any diagnosis should start.