20260115_112259_sol_wednesday_momentum.md

# AI Review: sol_wednesday_momentum **Date:** Thu Jan 15 11:22:59 CET 2026 **Strategy File:** /root/trade_rules/strategies/sol_wednesday_momentum.py --- ```json { "score": 45, "flags": ["overfitting", "inconsistent", "concentration"], "review": "## Strategy Assessment: sol_wednesday_momentum\n\n### Critical Issues\n\n**1. Day-of-Week Overfitting (Major)**\n\nThis strategy is built on a classic data-mining pattern: exploiting a specific day of the week. While the code uses \"standard\" parameters for EMAs, the core alpha depends entirely on Wednesday being special. Year-by-year results reveal this is unstable:\n\n- 2020: -23.8% (failed completely)\n- 2021: +105.6% (explosive)\n- 2022: +8.9% (barely positive, only 2 trades)\n- 2023: +90.8% (explosive again)\n- 2024: +54.0% (strong)\n- 2025: +29.2% (declining)\n\nThe strategy shows extreme regime dependency. It fails catastrophically in some years and produces outsized returns in others. This is textbook overfitting to calendar patterns that are known to be unstable in financial markets.\n\n**2. Inconsistent Year-Over-Year Performance**\n\nThe variance in annual returns is enormous:\n- Range: -23.8% to +105.6% (129% swing)\n- 2020 complete failure suggests the pattern doesn't hold in all market regimes\n- 2022 only had 2 trades, making it statistically insignificant\n- Returns are declining over time (105% → 90% → 54% → 29%), suggesting alpha decay\n\nA robust strategy should show reasonable consistency. This one depends heavily on which years you trade it.\n\n**3. Statistical Significance Concerns**\n\nSeveral years have very few trades:\n- 2020: 2 trades\n- 2022: 2 trades\n- 2025: 10 trades (partial year)\n\nWith so few observations, it's impossible to distinguish skill from luck. The 2020 result (0% win rate on 2 trades) is statistically meaningless but operationally painful.\n\n**4. Concentration Risk**\n\nWhile not explicitly shown, the explosive returns in 2021 (+105.6%) and 2023 (+90.8%) likely come from a handful of winning trades during bull markets. The strategy's success depends on catching these big moves, making it vulnerable to regime changes.\n\n### Positive Aspects\n\n- ✓ No lookahead bias - trades on next bar open\n- ✓ Uses standard EMA periods (20, 50, 200)\n- ✓ Clear regime filter prevents some bear market damage\n- ✓ Simple exit logic (time-based)\n- ✓ Passes momentum role validation gates\n- ✓ Code quality is good\n\n### Why This Scores Poorly\n\nThe fundamental problem is **calendar overfitting**. Trading based on day-of-week is a well-known pitfall. While the strategy dresses it up with EMA filters, the core edge depends on Wednesday being reliably stronger than other days. Academic research shows these patterns:\n\n1. Are often spurious (data mining artifacts)\n2. Decay rapidly when discovered\n3. Vary significantly by asset and time period\n4. Don't generalize to new market regimes\n\nThe year-by-year results confirm this: the pattern works in some years (2021, 2023) and fails in others (2020, 2022). A strategy that loses 23.8% in one year and makes 105.6% the next is not robust - it's regime-dependent and lucky.\n\n### Validation Performance Context\n\nThe validation return of 10.18% appears reasonable for momentum, but this doesn't overcome the fundamental overfitting concern. The strategy may have simply gotten lucky that the Wednesday pattern continued working in the validation period. Past calendar anomalies have disappeared entirely once widely known.\n\n### Recommendation\n\n**Score: 45/100 - Poor**\n\nThis strategy should be rejected or heavily revised. The day-of-week dependency is a red flag for overfitting, and the inconsistent year-over-year performance confirms it's not robust across market regimes. Even though it passes validation, the structural weakness makes it unsuitable for forward deployment.\n\nIf the researcher wants to pursue calendar-based strategies, they need to show the pattern is stable across multiple assets, time periods, and market conditions - not just curve-fit to one coin's historical quirks.", "rule_suggestion": "**Forbidden Pattern: Calendar-Based Strategies**\n\nStrategies that depend on specific days of the week (Monday, Wednesday, Friday, etc.), specific hours as primary alpha sources, or other calendar patterns are prohibited unless the researcher can demonstrate:\n\n1. The pattern holds across at least 3 different assets independently\n2. The pattern shows <30% variance in annual returns year-over-year\n3. A causal mechanism explains why the pattern should persist (not just \"historically strong\")\n4. The strategy works even when the calendar feature is randomized in sensitivity testing\n\nCalendar-based entry filters are acceptable as minor risk controls (e.g., \"avoid weekend gaps\") but should not be the primary source of alpha. Strategies where removing the calendar condition eliminates all edge are presumed overfit." } ```