20260115_070654_doge_ema_cascade_momentum.md

# AI Review: doge_ema_cascade_momentum **Date:** Thu Jan 15 07:06:54 CET 2026 **Strategy File:** /root/trade_rules/strategies_rejected/doge_ema_cascade_momentum.py --- ```json { "score": 25, "flags": ["overfitting", "inconsistent", "concentration"], "review": "## Critical Issues\n\n### 1. Severe Year-Over-Year Inconsistency\n\nThe strategy shows **catastrophic regime-specific behavior**:\n\n- **2021-2023**: Three consecutive years of losses (-47.3%, -43.2%, -11.1%)\n- **2024**: Sudden explosive gain (+243.9%) \n- **Post-2024**: Immediate reversion to near-zero performance (+4.9%)\n\nThis is a textbook example of **overfitting to a single regime**. The strategy clearly optimized for conditions that existed in one specific period and fails completely outside that window. The validation failure confirms this isn't generalizable.\n\n### 2. Dramatic Win Rate Collapse\n\nWin rates across years: 14.3% → 16.7% → 25.9% → 32.1% → 23.5%\n\nThe strategy only becomes \"profitable\" when win rate doubles from early years to the peak period, then immediately degrades. This suggests the entry/exit logic is **tuned to specific market microstructure** that existed temporarily.\n\n### 3. Concentration Risk (Likely)\n\nA +243.9% return in one year with only 28 trades and 32.1% win rate means approximately **9 winning trades generated 340%+ in gains** (accounting for the ~100% starting equity + 243% profit, minus losses from 19 losing trades).\n\nThis implies:\n- Top 3 trades likely account for >50% of total PnL\n- Strategy is highly dependent on catching a few extreme moves\n- One or two parameter changes would likely eliminate those specific entries\n\n### 4. Overfitting Red Flags\n\n**Despite claims of \"standard parameters,\" the logic is curve-fitted:**\n\n- \"Low touched within **2%** of EMA20 in last **5 bars**\" - these specific thresholds (2%, 5 bars) are not standard values from the rules\n- The **exact combination** of 4 EMAs (20/50/100/200) + pullback distance + lookback period creates a very narrow entry filter\n- 6 entry conditions (full cascade = 4 conditions, price > EMA20, recent pullback) at the complexity limit\n\n**The \"robustness testing\" claim is unverifiable** - stating \"tested SL 4-6%, pullback 3-10 bars\" doesn't help if those weren't the submitted parameters.\n\n### 5. False Regime Filter Claims\n\nThe strategy claims the EMA cascade \"stays flat during bear markets (54% of 2024-2025H1).\" But:\n\n- **The strategy took 14-18 trades per year in the worst performing years (2021-2022)**\n- If the filter truly worked, it should have **prevented most trades during those losing years**\n- The fact it still entered 14-18 positions and lost 40%+ means the filter is ineffective at its stated purpose\n\n### 6. Execution Concerns\n\nThe pullback detection uses `lows[j] <= ema20[j] * 1.02` which checks if low of **past bars** touched EMA20. While not same-candle execution, this creates a **very specific entry pattern** that may not repeat: it requires the exact sequence of (1) pullback happens, (2) EMA cascade still aligned, (3) price rebounds above EMA20, all within 5 bars.\n\n## Why This Fails\n\nThe core issue: **this strategy didn't discover a generalizable pattern—it reverse-engineered the entry conditions that would have captured specific historical moves.**\n\nThe EMA cascade + pullback timing + 2% threshold combination creates such a narrow filter that it only triggers on very specific price action patterns. When those patterns appeared (during one period), it won big. When they didn't (every other period), it failed.\n\n## Validation Results Confirm\n\n- **-18% return in validation** (worse than train average excluding the outlier year)\n- **Sharpe -1.17** (even worse than most train years)\n- Strategy completely broke down on unseen data\n\nThis is exactly what happens with overfit strategies: they excel in-sample on the regime they were tuned for, then collapse out-of-sample.\n\n## Verdict\n\n**REJECT** - This strategy should not be deployed. The year-over-year results show it's fundamentally regime-specific rather than robust. The single profitable year is an outlier, not evidence of edge.", "rule_suggestion": "**Consistency Gate:** Strategies must not show regime-specific outliers. If any single year's return exceeds the median annual return by more than 150 percentage points (e.g., +243% in one year vs ~0% median), AND that year contributes >70% of total cumulative return, flag for overfitting review. Require testing with that year excluded to verify the strategy has edge in other periods." } ```