The Regime Engine
How Warren turns a raw market into a labeled, graded, and honestly-projected regime read, computed only on closed candles.
Regime is the primitive the rest of Warren conditions on: strategy discovery ranks by regime fit, backtest windows are chosen by regime, position-size hints are discounted in weak regimes, and a directional call is withheld entirely when the regime cannot be read with confidence. This page explains how a raw market becomes a labeled, graded, trustworthy regime read, and how far forward that read can honestly be projected.
How the label is built
The regime engine is a weighted evidence aggregator. On each analyzed timeframe (by default 1h, 4h, and 1d), a family of well-known technical dimensions, trend, momentum, directional strength, volatility state, and volume or flow, each cast a vote toward bullish, bearish, or ranging. Higher timeframes use looser sensitivity than lower ones, because daily indicators move more slowly than hourly. Cross-timeframe agreement reinforces the aggregate; disagreement tilts the read toward ranging.
Derivatives (funding rate, open-interest change versus price) and market sentiment nudge the score as confirming context, and for altcoins broad-market context is folded in as a macro modifier. The result maps onto the seven-step conviction ladder, from strong bullish through sideways to strong bearish, with graded conviction.
- Closed candles only. The engine never scores a forming bar. An indicator computed on an unclosed candle "repaints" and effectively leaks the future, so closed-candle scoring makes every read reproducible by construction.
- Liquidity is information, not score. Order-book depth, positioning skew, and liquidation proxies are attached for trade management but deliberately excluded from the regime score, keeping the classification a clean function of price action (including volume), derivatives, and sentiment.
Hysteresis: refusing to flip-flop. A classifier recomputed every few hours would thrash on noise. The engine applies hysteresis at the combined and per-timeframe levels: changing the stored label requires conviction, a sufficient margin between the leading and runner-up reads plus corroboration from recent history, unless the old label is clearly losing, in which case the change is immediate. A change threshold wider than the revert threshold, applied to market state. A companion whipsaw read counts recent label flips and flags a fragile market even when the current label looks clean.
Two refinements shape the final label. A momentum read (strengthening, stable, or weakening) gates whether a "strong" label is allowed at all: a decaying trend cannot earn the strongest rung. And every read carries a single flat consensus block so consumers never have to triangulate across fields.
The multi-timeframe cascade
Each timeframe is classified independently, and the (1h, 4h, 1d) tuple maps to a cascade level describing how the timeframes relate. From it the engine derives a cascade master regime, the direction the timeframe structure supports with higher timeframes taking priority, which is often more trustworthy than the blended score. The read warns when the two disagree.
| Cascade level | Meaning | Confidence |
|---|---|---|
| Full cascade | All three timeframes agree on direction | high |
| Higher-TF trend | 4h and 1d agree; 1h counter-moving (a pullback, not a reversal) | medium-high |
| Partial cascade | 1h and 4h agree; the daily has not confirmed yet | medium |
| Daily only | Only the daily is directional; lower timeframes consolidating | low to medium |
| 1h diverging | Only the 1h is directional; higher timeframes ranging (fragile) | low |
| Conflicting / mixed | Directional but contradictory, or no coherent structure | low |
Why the cascade is central: a regime where all three timeframes agree tends to persist for days; one propped up by a single timeframe rarely survives past hours. Cross-timeframe alignment is the strongest available predictor of durability, so the cascade directly gates how far forward the engine will forecast.
Persistence and survival forecasting
Not a price prediction. This is the one place Warren forecasts forward, and it forecasts state duration, how long the current regime holds, never price direction. States have measurable lifetimes; prices do not have knowable paths.
The baseline is a cascade-gated persistence probability. Where enough history exists, it upgrades to an age-conditioned empirical survival estimate that accounts for how long the regime has already run: a regime weeks into its life has a different forward survival than one that started yesterday, which a memoryless model cannot express.
When the regime is too young or too few comparable spells exist, the forecast falls back or abstains rather than emitting a confident-looking number from thin data, the same discipline applied everywhere in Warren.
Choosing backtest windows from regime history
Historical regime windows are pre-computed per pair as dated spans, each labeled with its basic regime plus quality metadata: duration, the classification confidence it was labeled with, and the intra-window counter-trend drawdown it contains. The window planner searches these and ranks candidates quality-first, penalizing windows stitched from adjacent regimes. Tokens without their own history fall back through proxies, and every read reports which source it used:
When the best available window is weak, only stitched windows, low classification confidence, or heavy intra-window drawdown, the planner attaches explicit warnings instead of returning a clean-looking answer. A net-bullish window is not a monotonic uptrend, and that caveat is surfaced, not hidden.
Black box. The indicator set, per-timeframe thresholds, vote weights, derivatives and sentiment cutoffs, cascade multipliers, survival-support constants, and window-scoring weights are internal. The documented contract is the boundary: market data in, a graded label, cascade, persistence forecast, and coverage verdict out.