Regime Overlay Agent Skill

Regime Overlay — Triple-Confirmation Gate

A reusable filter that wraps any directional strategy and only allows entries when three independent regime confirmations align. Validated as the difference between a fragile and a robust trend strategy.

Searchable under: regime filter, trend gate, trade-or-skip filter, directional confirmation, ADX gate.

Why a separate primitive

After 21 backtests, the single biggest performance lever was not entry logic, stop sizing, or trailing config. It was whether the strategy traded at all during specific market structure.

The proof:

• Same Donchian-breakdown short on BTC 4h, no regime gate → +12.85% over 162d (looks great)
• Same strategy, on a different window (the rangy second half) → -4.32% (edge gone)
• Same strategy with triple-confirmed regime gate → +6.69% with 0% drawdown, fires zero trades in the chop window

The first version captured a real edge mixed with random noise. The third version captured only the real edge. Smaller absolute return; cleaner equity curve; no losing periods.

The three confirmations

A signal is "in regime" when ALL three of these hold simultaneously:

| Indicator | Threshold | What it confirms |

|---|---|---|

| (EMA50 - EMA200) / EMA200 | < -0.06 (for bear) | Structural trend direction — not a fresh cross, real separation |

| ADX(14) | > 25 | Strength of the current trend |

| close.pct_change(30) | < -0.10 (for bear) | Recent momentum is actually moving |

Three independent angles — structure, strength, recency. A market can satisfy any one or two by accident; satisfying all three is genuinely rare and genuinely informative.

For long-side strategies, invert the signs:

• (EMA50 - EMA200) / EMA200 > +0.06
• ADX(14) > 25 (same — ADX is direction-agnostic)
• close.pct_change(30) > +0.10

Implementation

def regime_strong_bear(df: pd.DataFrame) -> pd.Series:
    ema50 = ta.EMA(df, timeperiod=50)
    ema200 = ta.EMA(df, timeperiod=200)
    adx = ta.ADX(df, timeperiod=14)
    ema_sep = (ema50 - ema200) / ema200
    ret_30 = df["close"].pct_change(30)
    return (ema_sep < -0.06) & (adx > 25) & (ret_30 < -0.10)


def regime_strong_bull(df: pd.DataFrame) -> pd.Series:
    ema50 = ta.EMA(df, timeperiod=50)
    ema200 = ta.EMA(df, timeperiod=200)
    adx = ta.ADX(df, timeperiod=14)
    ema_sep = (ema50 - ema200) / ema200
    ret_30 = df["close"].pct_change(30)
    return (ema_sep > 0.06) & (adx > 25) & (ret_30 > 0.10)


def regime_range(df: pd.DataFrame) -> pd.Series:
    adx = ta.ADX(df, timeperiod=14)
    return adx < 25

The range-regime check is the opposite end of the spectrum — useful for gating mean-reversion strategies (see bollinger-reverter-4h).

Expected behavior

Apply this overlay to any signal-driven entry. The cost is fewer trades; the benefit is the trades you do take fire in the right environment.

| Strategy | Without overlay | With overlay |

|---|---|---|

| Donchian breakdown BTC | 23 trades, 56.5% win, +12.85%, 3.69% DD | 6 trades, 100% win, +6.69%, 0% DD |

| (Same, second-half only) | 8 trades, 25% win, -4.32% | 0 trades, no PnL |

The overlay traded the second column's win rate and drawdown for the first column's absolute return. Most production strategies should prefer the second column.

Tunables

| Parameter | Conservative | Aggressive | Notes |

|---|---|---|---|

| EMA separation threshold | 0.08 | 0.04 | Tighter = wait for deeper structural moves |

| ADX threshold | 30 | 20 | Higher = stricter trend-strength requirement |

| Return lookback bars | 40 | 20 | Window for "actual momentum" check |

| Return threshold | 0.15 | 0.06 | How much momentum needed |

Start conservative (0.06 / 25 / -0.10). Loosen only if the gate eliminates so many trades that the strategy's edge can't compound.

What this is NOT

• Not a signal in itself. The overlay only says "regime is right"; it doesn't tell you when to enter
• Not magic. In an environment where the historical regime gate would never fire (e.g., a yearlong sideways grind), the strategy simply stays out. That's the feature
• Not a replacement for stops or take-profit. Even regime-confirmed signals fail individually

When NOT to use a regime gate

• Mean-reversion strategies (bollinger-reverter-4h) work IN range regimes — use regime_range (ADX < 25) instead of bear/bull strong gates
• Funding-carry strategies that earn rebates rather than directional moves
• Pair-arbitrage or basis-arb strategies where the edge is structural, not regime-conditional

Pairing patterns

Most directional strategies in this repo benefit from this overlay:

• donchian-strong-regime — uses bear overlay (validated)
• Any custom trend-follow strategy

The overlay file itself is regime-agnostic — it provides functions for both bear and bull, and a range complement.