Most traders assume a strategy that stops making money must contain a bug, a broken data feed, or a mistake in execution. Usually none of that is true. The code runs exactly as written, the signals fire on schedule, and the model still produces confident predictions. It has simply started predicting a market that no longer exists. This is model decay, and it is the most common way that a genuinely good trading system quietly turns into a losing one.
The model did not break — the market moved
A trading model is a compressed description of past behaviour. It has learned which combinations of price, volatility, order flow and macro context tended to precede a move. As long as tomorrow resembles the period it was trained on, that description stays useful. The problem is that the thing being described is not a fixed object. Markets are made of people, institutions and other algorithms adapting to each other, and their collective behaviour shifts. When it shifts far enough, the model is still answering the old question accurately while the market has started asking a new one.
Stationarity: the assumption every model quietly makes
Every statistical model assumes some degree of stationarity — the idea that the relationship between inputs and outcomes stays roughly constant over time. Financial series break this assumption more aggressively than almost any other domain. The average return, the volatility, and crucially the correlations between assets all drift and sometimes lurch. A model fitted to a calm, trending regime encodes relationships that were real then and are simply absent now. Nothing in the mathematics warns you, because the model was never told that the ground could move. It extrapolates with the same confidence whether the world has changed or not.
Where regime change actually comes from
Regime change is not mystical. It has identifiable sources. A shift in central-bank posture rewires the relationship between interest rates and every risk asset. A change in who is trading — retail flows receding, a large systematic fund entering — alters the microstructure that short-horizon models feed on. Liquidity conditions expand and contract, changing how far price travels on a given order. And participants adapt to each other: once an edge is widely known, the very act of trading it erodes the pattern that made it profitable. Each of these changes the data-generating process, which is the technical way of saying the rules that produce the numbers your model learned from.
Concept drift is not overfitting
It is worth separating two failures that look alike from the outside. Overfitting is a training-time problem: the model memorised noise in the historical sample and never had real predictive power to begin with. Concept drift is a deployment-time problem: the model had genuine power, and then the underlying relationship changed and took that power away. The distinction matters because the remedies differ. Overfitting is fixed with better validation and simpler models before you ever go live, a theme explored in this piece on models that learn the noise. Drift cannot be fixed in advance at all; it can only be monitored for and adapted to after deployment.
Why decay is silent
The dangerous feature of model decay is that profit and loss is a lagging alarm. By the time the equity curve has visibly rolled over, the model has often been trading a dead edge for weeks. Returns are noisy, so a run of losses is indistinguishable from ordinary variance until the sample is large enough to be sure — and by then the damage is done. Worse, a decaying model frequently becomes more confident, not less, because the new regime can produce input patterns that sit at the extremes of what it saw in training. The system is most emphatic precisely when it is most wrong.
Detecting drift before the drawdown
The way out is to stop watching only the P&L and start watching the inputs. If the statistical distribution of the features the model consumes moves away from the distribution it was trained on, that is an early warning that does not depend on losing money first. The same applies to the model's own error pattern: rising prediction error, or errors that suddenly cluster in one direction, signals that the relationship has shifted before the account balance confirms it. None of this tells you what the new regime is. It tells you, usefully and early, that the old one has ended and the model's assumptions are no longer being met.
What would prove this wrong
This view is falsifiable, which is what makes it worth holding. If trading models retained their edge indefinitely without retraining, and if strategies published years ago continued to perform as originally reported, the decay thesis would be wrong and markets would be closer to stationary than claimed. If input distributions stayed stable across policy shifts and liquidity cycles, monitoring them would be pointless. The reason experienced desks retrain, re-validate and retire models on a schedule is that the opposite is repeatedly observed: edges fade, distributions move, and the models that survive are the ones built to expect it.
Building for decay, not against it
The practical shift is to treat every model as perishable rather than permanent. That means shorter deployment horizons, explicit monitoring of input and error distributions, and a plan to reduce size or stand aside when the market stops resembling the training data. A slightly less powerful model that you can trust to fail loudly is more valuable than a strong one that fails in silence. The goal is not a system that never decays — that does not exist — but one that tells you it is decaying while you can still act on the warning.
This article is for general market information and education. It does not constitute financial or investment advice. Trading carries risk, and past performance is not indicative of future results.
0 Comments