Why Regime-Based Backtesting

This is the single most important idea in how Warren evaluates strategies. A single continuous backtest averages bull, bear, and ranging behavior into one number that hides both when a strategy works and when it quietly gives everything back.

Backtest a trend follower continuously over a long window and you get one figure, say +9%, that is a weighted blend of three completely different behaviors: large gains while the trend held, large losses while it inverted, and slow bleed from fees and whipsaw while the market ranged. That one number is not wrong, exactly. It is just useless for the decision you actually have to make.

The averaging trap

The blended number tells you almost nothing actionable. It does not tell you the strategy is a trend strategy; it does not tell you to turn it off in a downtrend. Two strategies with identical full-period returns can have opposite regime profiles: one grinds steadily, the other makes 40% in Q1 and loses 31% in Q3. Averaged, they look the same. Deployed, they are not.

The same trap operates across instruments: a strategy that loses 30% on one alt while winning 15% on a major shows a mediocre blended average that hides a real per-pair winner. Averaging hides the conditional structure of edge, which is exactly the structure you need to deploy and manage capital.

The +9% and the segmented profile describe the same trades. Only the second one tells you what to do with them.

Two modes, chosen by the question

Warren labels every backtest with the question it can honestly answer, selecting the mode automatically from the window's length (always overridable).

The two views are complementary, not competing. The full-period view is the honest reality check on the regime view: it prices in the cost of not switching. The regime view is the honest reality check on the full-period view: it decomposes a bland average into "great here, disastrous there," which is what you need to deploy and manage capital.

Robustness as a first-class action

Before a regime backtest, Warren's window planner searches pre-computed regime history for a clean, confidently-classified, single-regime window. Candidates are ranked quality-first, not recency-first, because an older but unambiguous window is a stronger test environment than a recent noisy one.

Its flagship use is the all-regimes robustness sweep: one request returns the best window for each of bullish, bearish, and ranging, the same strategy is run across all three, and it is judged on its worst regime, not its favorite.

The regime-conditional thesis is also encoded as a prior. Each strategy category maps to the regimes where it should have an edge, and the planner nudges testing toward the regime where the category's edge, if any, lives.

A strategy that performs "the same everywhere" is either genuinely all-weather (rare, and usually defensive by construction) or, far more commonly, its edge is too small to distinguish from noise in any regime.

Honest trade-offs

Regime-based testing is the right default, but Warren surfaces its limits rather than papering over them:

• A regime-scoped test read in isolation flatters the strategy. Mitigated by sweeping all three regimes and grading the worst.
• Regime labels are probabilistic, and a net-directional window can still contain counter-trend stretches. These are flagged, not hidden.
• Tokens without their own regime history fall back to market proxies. This is reported, with caveats.
• Single-regime windows contain fewer trades, so per-regime statistics carry wider error bars.
• The regime-conditional edge is only realizable if you correctly detect the regime live and switch in time, a cost the full-period view captures.
• No mitigation is left to prose alone: the robustness grade operationalizes the lesson with per-regime grades, a consistency score (worst divided by best), and a hard disqualification on catastrophic drawdown in any single regime. A great average cannot buy an A.

The recommended evaluation flow

The output of this flow is not a single +9%. It is a profile: earns in trends, bleeds in ranges, disqualified in bear markets by drawdown, deploy now because we are in a confirmed uptrend, switch out when the regime turns. That profile is what a continuous backtest can never give you.