Regime Filters — When a Trading Model Should Sit Out

A regime filter in a trading model is a separate logic layer that decides whether the current market state is one in which the model is allowed to trade. The entry logic might say "this is a setup"; the regime filter says "yes, but only if the market is in a state where that setup historically works." Together, they cut the model's exposure to conditions where its edge is unlikely to hold.

This matters because most trading models do not fail through bad entries. They fail because the market they were designed for has temporarily turned into a different market. A trend-following model in a sideways range will produce textbook entries and lose on most of them. A mean-reversion model during a strong trend will do the same. Regime filters are the design choice that prevents the model from playing in the wrong conditions in the first place.

What the regime filter does that the entry logic cannot

Entry logic identifies a pattern — a breakout, a pullback, a cross, a level touch. The pattern exists or it doesn't. What the entry logic cannot easily encode is whether the wider market context makes that pattern reliable right now.

A pullback entry signal in a healthy trend is a high-probability setup. The same pullback signal in a choppy, directionless market is a coin flip. The signal itself looks identical in both cases — the difference is the regime. The regime filter is the place in the model where that distinction is made explicit.

In darwintIQ, the regime filter is one of three components of every trading model, alongside the Entry Logic and the Position Manager. The Genetic Algorithm evolves the combination, not the components in isolation, which is why a model's entry logic and its regime filter tend to develop a meaningful relationship over generations: filters are selected to admit the conditions where the entry logic actually works.

The filter types darwintIQ models can use

The available options each describe market state in a different way.

SMA filters define regime by the relationship between price and a simple moving average. They are the most common form of trend filter — typically requiring price to be above a longer-period SMA before long trades are taken, and below it for shorts. The SMA filter is blunt but effective: it cleanly removes a model from countertrend exposure.

RSI filters use the Relative Strength Index to classify the market as overbought, oversold, or neutral. A momentum entry might be allowed only when RSI confirms direction; a mean-reversion entry only when RSI has reached an extreme. RSI filters are more responsive than SMA filters but more prone to misclassifying short-term spikes.

RsiBand filters define an acceptable RSI corridor — for example, between 40 and 60 — and allow trades only when momentum is neither extreme nor exhausted. RsiBand is well-suited to range-bound entry logic that depends on the market not having committed to a direction.

SupResFilter uses recent support and resistance structure to decide whether the market is in a respect-the-levels regime or in a break-the-levels regime. Models that depend on structural levels — the SupRes position manager is the natural companion — benefit from a filter that recognises when those levels are being honoured.

TrendRegimeFilter uses a composite measure of directional persistence, not just price-to-MA distance. A market can be above its SMA without trending in any useful sense; TrendRegimeFilter is designed to distinguish genuine trends from drift.

NoFilter is the explicit absence of a regime filter. A model running NoFilter is permitted to take every entry signal regardless of conditions. This is sometimes the right design choice — frequent signals from a robust entry logic can outperform the same logic strangled by an over-restrictive filter — but it asks more of the entry logic and the position manager.

When a regime filter helps and when it hurts

The case for a regime filter is empirical. If a model's losing trades cluster in a recognisable market state — sideways periods for a trend follower, strong directional moves for a mean reverter — then a filter that excludes that state will improve risk-adjusted return. The model takes fewer trades, but the ones it takes have a better expected value, and aggregate metrics like Profit Factor and Calmar improve as a result.

The cost is sample size. Every filter reduces the number of trades the model takes. Below a certain frequency, performance metrics become statistically thin and small changes in market conditions can dominate the recent record. A filter that improves trade quality but cuts trade count by 80% may produce a better-looking strategy that is, in practice, harder to evaluate and slower to recover from drawdown.

The second cost is regime classification error. Every filter is itself a model of market state, and that model can be wrong. An SMA filter does not know that a trend has just ended; a TrendRegimeFilter can be slow to recognise a new directional move. When the filter misclassifies the regime, the model either misses real opportunities or, worse, takes trades in the conditions it was designed to avoid. This is why edge decay often shows up in the regime filter before it shows up anywhere else.

The right way to evaluate a filter is to compare the same entry logic and position manager with and without it across multiple windows. If the filter consistently improves risk-adjusted return without crushing trade count, it is doing its job. If it improves the headline metrics on the in-sample window but not the out-of-sample one, the filter has been fitted to a regime that is no longer present.

How darwintIQ selects regime filters in practice

Within the darwintIQ population, models with explicit regime filters tend to dominate during stable conditions, because their filtering reduces the variance of their results. Models running NoFilter tend to do better through regime transitions, because they continue to trade rather than waiting for the filter to confirm a regime that the market has only just entered.

The Genetic Algorithm does not choose between these styles abstractly. It rewards whichever combination is producing better results on the rolling 4-hour evaluation window. As conditions change, the dominant filter types in the leaderboard change with them — which is precisely the design intent. The regime filter is not a setting; it is a hypothesis about the current market that the model is making in real time.

Final thoughts

Regime filters are one of the few parts of a trading model that explicitly acknowledge what kind of market the model was built for. Used well, they sharpen the entry logic by removing the conditions that produce its weakest trades. Used badly, they overfit to a regime that has already passed and produce a model that looks elegant on history and freezes in the present. The discipline is the same as everywhere else in quantitative trading: design honestly, validate out-of-sample, and let the data — not the elegance of the idea — decide whether the filter belongs in the model.