Why trading strategies stop working

By John, builder of Wise Apple · Updated August 27, 2026

It worked. Not in a backtest — live, for five months, close to what the test predicted. Then it went flat, then it went down, and you are staring at the chart trying to decide whether the market changed or you were lucky.

"The regime changed" is the most comforting diagnosis in trading, because it means the strategy was right and the world moved. It is sometimes true. But four other things produce exactly the same experience, and three of them mean the strategy was never what you thought. Telling them apart is the difference between waiting out a drawdown and funding one.

This guide is about the five explanations for a strategy that stops working, the fingerprint each one leaves, and the honest response to each.

First, rule out the boring answer

Before regime, before crowding, before anything interesting: is the bad stretch inside what the backtest predicted? Every strategy has losing streaks, and a backtest with honest drawdown numbers tells you how long and how deep they get. If your worst backtested drawdown was 22% over seven months and you are four months and 14% into a live one, nothing has happened yet. The strategy is doing what it said it would do, and the only thing being tested is you.

Most strategies are abandoned inside a drawdown the backtest forecast. Read the drawdown numbers before you read anything else, and if you did not believe them when you started, that was the mistake.

The five explanations

1. Variance

Covered above, but it deserves its own name. A strategy with a 55% win rate will produce runs of eight losses in a row with unremarkable regularity. Five months of good results followed by two bad ones is within the ordinary noise of almost any strategy. The fingerprint: the live results are inside the backtest's drawdown envelope, the trades still look like the backtest's trades, and nothing about the market is obviously different.

2. Overfitting that took a while to show

Overfit strategies usually fail immediately, but not always. If the live period happened to resemble the tuned period for a few months — same trend, same volatility — an overfit strategy can look fine until the resemblance ends. The fingerprint: performance was never robust to parameter changes even in the backtest, and the good live months coincided with conditions that closely matched the in-sample data. The overfitting guide has the tests.

3. Crowding

If enough capital trades the same signal, the signal stops paying. The edge was real, and it was consumed. This is the fate of most published anomalies and many popular indicator strategies. The fingerprint is gradual: returns decay over quarters rather than collapsing, slippage on entries gets worse as other participants front the same moves, and the strategy's trades increasingly coincide with sharp reversals — because everyone else got in at the same moment.

4. Cost creep

The market did not change; the bill did. Spreads widened, funding turned against you, your size grew into the impact zone, the exchange changed its fee tier. Strategies with thin edges are especially vulnerable, and the fingerprint is that gross results look fine while net results drift down. Rerun the backtest at current costs and see whether the edge is still there; the costs guide covers what to charge.

5. Regime change

Now the real thing. Markets have modes — trending and ranging, high volatility and low, risk-on and risk-off — and a strategy built for one mode does not work in another. Trend-following strategies bleed in ranges; mean-reversion strategies get run over in trends. When the mode shifts, a strategy tuned to the old mode stops earning, and it does so abruptly.

The fingerprint: something measurable about the market is different. Realised volatility doubled or halved. The average trend length changed. Correlations that used to hold broke. And crucially, the backtest — if it covered enough history — contains earlier stretches with the same measurable difference where the strategy also struggled.

The early-warning system

You do not have to wait for the drawdown to find out. Rolling out-of-sample evaluation — re-testing the frozen rules on each new window as it closes — shows decay while it is still small. Compare the most recent window to the distribution of earlier windows. A recent window at the bottom of that distribution is variance. Three in a row at the bottom is a trend, and the question becomes which of the five it is.

This is what walk-forward analysis gives you when you keep running it after the strategy goes live: a continuously updated answer to "is this still working", with error bars, rather than a feeling.

Building for regimes from the start

If regime change is the diagnosis, the strategy does not need fixing. It needs to know when to stop. Three approaches, in increasing complexity:

1. A volatility or trend filter. Only take trades when the market is in the mode the strategy was built for. A simple ADX or realised-volatility gate removes a surprising share of the damage, at the cost of fewer trades.
2. Separate strategies per regime. A trend system and a range system, each with its own filter, so that one is usually on duty.
3. A regime feature in the model. If you use machine learning, feed it a regime measure as an input and let it learn that the same setup means different things in different modes. This is harder to get right than a filter and easier to overfit.

Whatever you choose, backtest it across every regime the history contains, and count the trades in each. A regime filter that has only ever been tested in one regime is a new way to be wrong.

How this looks in Wise Apple

Wise Apple ships regime-aware inputs as ordinary settings. The Volatility, ADX and HMM entry filters gate trades on measurable market state, and the feature set includes an HMM regime feature and trend-quality measures (efficiency ratio, choppiness, Hurst) that a model can learn from. The Walk-Forward Window reports each out-of-sample period separately, which is the early-warning view above, and the alert node keeps evaluating the frozen strategy on new bars so the comparison between "what the backtest said" and "what is happening" is always current. None of it predicts the next regime. It tells you which one you are in.