Antigravity ML Crypto Predictor

Strategy: ICT Fair Value Gap (Selective) — Same idea as the Fair Value Gap strategy but more picky — it only triggers when multiple extra conditions are met (stron

Signal reason: ICT FVG discount zone, ADX=65, RSI=35

MFE: +1.6% (best price reached) | MAE: -3.5% (worst price reached)

❌ What went wrong:

• Stop loss hit within 24 hours — entered against a strong trend continuation
• Adverse move (-3.5%) was 2.1× larger than favorable move (1.6%) — price moved strongly against us
• Volume ratio was only 0.60× average — weak conviction, big players weren't participating
• 3 losses on ETH-USD — asset was in a sustained downtrend that multiple strategies failed to detect. This means the per-strategy filters weren't enough to catch correlation risk.
• The Fair Value Gap was not respected by institutional buyers — the expected 'fill' didn't produce a bounce. This often happens when the broader trend overwhelms the FVG zone.

🔧 Proposed tweaks:

• Increase SL width by 0.5× ATR or add a trend-strength filter (ADX > 40 = skip)
• Add a minimum R:R filter: only trade when ATR-adjusted TP distance > 2× SL distance
• Require volume_ratio > 1.0 for entry (strong volume confirmation)
• Add a correlation guard: if 2+ strategies lose on ETH-USD within 48h, block new entries for 72h
• Add higher-timeframe (4h) trend confirmation — only enter FVG longs when 4h EMA20 > EMA50

🤖 Self-learning status:

• ✅ AUTO-FIX: The nightly retrain will see this SL_HIT in the training data. Over time, the model learns to avoid entries where rapid SL hits occurred under similar conditions (same RSI range, same trend direction, same volatility regime).
• ✅ AUTO-FIX: The model feeds MFE/MAE ratios back into training features. After seeing enough trades where MAE >> MFE, it learns to weight entry conditions that produce BETTER MFE/MAE ratios.
• ⚠️ PARTIAL: The self-learning loop can learn that low-volume entries correlate with losses, but the volume_ratio threshold itself needs a MANUAL adjustment in the strategy config. We've flagged this for the next code update.
• ❌ CANNOT SELF-FIX: The self-learning loop operates per-strategy, not cross-strategy. A correlation guard between strategies requires a NEW CODE MODULE that monitors active positions across all strategies. This is a feature gap.
• ⚠️ PARTIAL: The model can learn FVG entries in downtrends fail, but the FVG strategy itself doesn't check higher timeframe trends. Needs a MANUAL multi-timeframe filter addition.

📋 What else is needed:

• ⏳ Need 5+ more SL_HIT samples from this strategy to build a statistically reliable 'SL too tight' pattern. Currently at 3/5 SL hits.
• 🔧 MANUAL CHANGE NEEDED: Update the alpha_engine strategy filters to require min_volume_ratio=1.0. This cannot be learned automatically — it's a hard config change.
• 🆕 FEATURE NEEDED: Build a CorrelationGuard module in the alpha_engine that tracks open positions per symbol across all strategies and enforces max_positions_per_asset=1.
• 🔧 MANUAL CHANGE: Add a 4h trend filter to FVG strategies: if 4h EMA20 < EMA50, block all BUY FVG signals. This is the #1 fix needed.

Strategy: ICT Fair Value Gap (Selective) — Same idea as the Fair Value Gap strategy but more picky — it only triggers when multiple extra conditions are met (stron

Signal reason: ICT FVG discount zone, ADX=48, RSI=45

MFE: +0.8% (best price reached) | MAE: -4.1% (worst price reached)

❌ What went wrong:

• Stop loss hit within 24 hours — entered against a strong trend continuation
• Volume ratio was only 0.58× average — weak conviction, big players weren't participating
• RSI was 44 at entry — FVG strategies work best when RSI < 35 (deeply oversold). At 44, price wasn't actually at a discount.
• 2 losses on DOGE-USD — asset was in a sustained downtrend that multiple strategies failed to detect. This means the per-strategy filters weren't enough to catch correlation risk.
• The Fair Value Gap was not respected by institutional buyers — the expected 'fill' didn't produce a bounce. This often happens when the broader trend overwhelms the FVG zone.

🔧 Proposed tweaks:

• Increase SL width by 0.5× ATR or add a trend-strength filter (ADX > 40 = skip)
• Require volume_ratio > 1.0 for entry (strong volume confirmation)
• Tighten RSI filter to < 35 for FVG-based strategies
• Add a correlation guard: if 2+ strategies lose on DOGE-USD within 48h, block new entries for 72h
• Add higher-timeframe (4h) trend confirmation — only enter FVG longs when 4h EMA20 > EMA50

🤖 Self-learning status:

• ✅ AUTO-FIX: The nightly retrain will see this SL_HIT in the training data. Over time, the model learns to avoid entries where rapid SL hits occurred under similar conditions (same RSI range, same trend direction, same volatility regime).
• ⚠️ PARTIAL: The self-learning loop can learn that low-volume entries correlate with losses, but the volume_ratio threshold itself needs a MANUAL adjustment in the strategy config. We've flagged this for the next code update.
• ⚠️ PARTIAL: Model sees the RSI value as a feature and can learn to avoid entries at RSI > 40, but the FVG strategy's RSI threshold is a HARD FILTER that needs manual adjustment in the strategy code.
• ❌ CANNOT SELF-FIX: The self-learning loop operates per-strategy, not cross-strategy. A correlation guard between strategies requires a NEW CODE MODULE that monitors active positions across all strategies. This is a feature gap.
• ⚠️ PARTIAL: The model can learn FVG entries in downtrends fail, but the FVG strategy itself doesn't check higher timeframe trends. Needs a MANUAL multi-timeframe filter addition.

📋 What else is needed:

• ⏳ Need 5+ more SL_HIT samples from this strategy to build a statistically reliable 'SL too tight' pattern. Currently at 3/5 SL hits.
• 🔧 MANUAL CHANGE NEEDED: Update the alpha_engine strategy filters to require min_volume_ratio=1.0. This cannot be learned automatically — it's a hard config change.
• 🔧 MANUAL CHANGE NEEDED: In the FVG strategy config, change rsi_threshold from 50 to 35. Self-learning can't change strategy parameters — only entry probabilities.
• 🆕 FEATURE NEEDED: Build a CorrelationGuard module in the alpha_engine that tracks open positions per symbol across all strategies and enforces max_positions_per_asset=1.
• 🔧 MANUAL CHANGE: Add a 4h trend filter to FVG strategies: if 4h EMA20 < EMA50, block all BUY FVG signals. This is the #1 fix needed.

Strategy: Smart Money Fair Value Gap — Tracks where big institutional traders (banks, hedge funds) are likely buying. When price drops into a zone where instit

Signal reason: Bullish FVG fill zone (1999.50-2042.64), ADX=65, RSI=35

MFE: +1.6% (best price reached) | MAE: -4.1% (worst price reached)

❌ What went wrong:

• Adverse move (-4.1%) was 2.6× larger than favorable move (1.6%) — price moved strongly against us
• 3 losses on ETH-USD — asset was in a sustained downtrend that multiple strategies failed to detect. This means the per-strategy filters weren't enough to catch correlation risk.
• The Fair Value Gap was not respected by institutional buyers — the expected 'fill' didn't produce a bounce. This often happens when the broader trend overwhelms the FVG zone.

🔧 Proposed tweaks:

• Add a minimum R:R filter: only trade when ATR-adjusted TP distance > 2× SL distance
• Add a correlation guard: if 2+ strategies lose on ETH-USD within 48h, block new entries for 72h
• Add higher-timeframe (4h) trend confirmation — only enter FVG longs when 4h EMA20 > EMA50

🤖 Self-learning status:

• ✅ AUTO-FIX: The model feeds MFE/MAE ratios back into training features. After seeing enough trades where MAE >> MFE, it learns to weight entry conditions that produce BETTER MFE/MAE ratios.
• ❌ CANNOT SELF-FIX: The self-learning loop operates per-strategy, not cross-strategy. A correlation guard between strategies requires a NEW CODE MODULE that monitors active positions across all strategies. This is a feature gap.
• ⚠️ PARTIAL: The model can learn FVG entries in downtrends fail, but the FVG strategy itself doesn't check higher timeframe trends. Needs a MANUAL multi-timeframe filter addition.

📋 What else is needed:

• 🆕 FEATURE NEEDED: Build a CorrelationGuard module in the alpha_engine that tracks open positions per symbol across all strategies and enforces max_positions_per_asset=1.
• 🔧 MANUAL CHANGE: Add a 4h trend filter to FVG strategies: if 4h EMA20 < EMA50, block all BUY FVG signals. This is the #1 fix needed.

Strategy: Altcoin Season Rotation — When Bitcoin goes up, money eventually flows into smaller coins (altcoins). This strategy detects when that 'rotation' i

Signal reason: SOL outperforms BTC by 4.1% (7d), BTC dominance=56.1%, phase=late euphoria (22mo post-halving). Liu & Tsyvinski JF 2021: crypto momentum factor

MFE: +1.7% (best price reached) | MAE: -5.5% (worst price reached)

❌ What went wrong:

• Stop loss hit within 24 hours — entered against a strong trend continuation
• Adverse move (-5.5%) was 3.2× larger than favorable move (1.7%) — price moved strongly against us

🔧 Proposed tweaks:

• Increase SL width by 0.5× ATR or add a trend-strength filter (ADX > 40 = skip)
• Add a minimum R:R filter: only trade when ATR-adjusted TP distance > 2× SL distance

🤖 Self-learning status:

• ✅ AUTO-FIX: The nightly retrain will see this SL_HIT in the training data. Over time, the model learns to avoid entries where rapid SL hits occurred under similar conditions (same RSI range, same trend direction, same volatility regime).
• ✅ AUTO-FIX: The model feeds MFE/MAE ratios back into training features. After seeing enough trades where MAE >> MFE, it learns to weight entry conditions that produce BETTER MFE/MAE ratios.

📋 What else is needed:

• ⏳ Need 5+ more SL_HIT samples from this strategy to build a statistically reliable 'SL too tight' pattern. Currently at 1/5 SL hits.

Strategy: MVRV Ratio Proxy — Compares the current market value to the 'realized value' (what people actually paid). When MVRV is low, crypto is 'chea

Signal reason: MVRV proxy 0.57 (below realized), Z-score -1.4, RSI 36 recovering

MFE: +0.5% (best price reached) | MAE: -5.1% (worst price reached)

❌ What went wrong:

• Stop loss hit within 24 hours — entered against a strong trend continuation
• 3 losses on ETH-USD — asset was in a sustained downtrend that multiple strategies failed to detect. This means the per-strategy filters weren't enough to catch correlation risk.

🔧 Proposed tweaks:

• Increase SL width by 0.5× ATR or add a trend-strength filter (ADX > 40 = skip)
• Add a correlation guard: if 2+ strategies lose on ETH-USD within 48h, block new entries for 72h

🤖 Self-learning status:

• ✅ AUTO-FIX: The nightly retrain will see this SL_HIT in the training data. Over time, the model learns to avoid entries where rapid SL hits occurred under similar conditions (same RSI range, same trend direction, same volatility regime).
• ❌ CANNOT SELF-FIX: The self-learning loop operates per-strategy, not cross-strategy. A correlation guard between strategies requires a NEW CODE MODULE that monitors active positions across all strategies. This is a feature gap.

📋 What else is needed:

• ⏳ Need 5+ more SL_HIT samples from this strategy to build a statistically reliable 'SL too tight' pattern. Currently at 1/5 SL hits.
• 🆕 FEATURE NEEDED: Build a CorrelationGuard module in the alpha_engine that tracks open positions per symbol across all strategies and enforces max_positions_per_asset=1.

Strategy: ICT Fair Value Gap (Selective) — Same idea as the Fair Value Gap strategy but more picky — it only triggers when multiple extra conditions are met (stron

Signal reason: ICT FVG discount zone, ADX=69, RSI=35

MFE: +0.1% (best price reached) | MAE: -1.9% (worst price reached)

❌ What went wrong:

• Stop loss hit within 24 hours — entered against a strong trend continuation
• Volume ratio was only 0.60× average — weak conviction, big players weren't participating
• The Fair Value Gap was not respected by institutional buyers — the expected 'fill' didn't produce a bounce. This often happens when the broader trend overwhelms the FVG zone.

🔧 Proposed tweaks:

• Increase SL width by 0.5× ATR or add a trend-strength filter (ADX > 40 = skip)
• Require volume_ratio > 1.0 for entry (strong volume confirmation)
• Add higher-timeframe (4h) trend confirmation — only enter FVG longs when 4h EMA20 > EMA50

🤖 Self-learning status:

• ✅ AUTO-FIX: The nightly retrain will see this SL_HIT in the training data. Over time, the model learns to avoid entries where rapid SL hits occurred under similar conditions (same RSI range, same trend direction, same volatility regime).
• ⚠️ PARTIAL: The self-learning loop can learn that low-volume entries correlate with losses, but the volume_ratio threshold itself needs a MANUAL adjustment in the strategy config. We've flagged this for the next code update.
• ⚠️ PARTIAL: The model can learn FVG entries in downtrends fail, but the FVG strategy itself doesn't check higher timeframe trends. Needs a MANUAL multi-timeframe filter addition.

📋 What else is needed:

• ⏳ Need 5+ more SL_HIT samples from this strategy to build a statistically reliable 'SL too tight' pattern. Currently at 3/5 SL hits.
• 🔧 MANUAL CHANGE NEEDED: Update the alpha_engine strategy filters to require min_volume_ratio=1.0. This cannot be learned automatically — it's a hard config change.
• 🔧 MANUAL CHANGE: Add a 4h trend filter to FVG strategies: if 4h EMA20 < EMA50, block all BUY FVG signals. This is the #1 fix needed.

Strategy: Smart Money Fair Value Gap — Tracks where big institutional traders (banks, hedge funds) are likely buying. When price drops into a zone where instit

Signal reason: Not logged

MFE: +2.1% (best price reached) | MAE: -5.6% (worst price reached)

❌ What went wrong:

• Adverse move (-5.6%) was 2.7× larger than favorable move (2.1%) — price moved strongly against us
• RSI was 50 at entry — FVG strategies work best when RSI < 35 (deeply oversold). At 50, price wasn't actually at a discount.
• Trade expired without hitting TP or SL — no clear directional move materialized. The model was right about no crash, but wrong about an upward move.

🔧 Proposed tweaks:

• Add a minimum R:R filter: only trade when ATR-adjusted TP distance > 2× SL distance
• Tighten RSI filter to < 35 for FVG-based strategies
• Reduce max hold time from 72h to 48h; add exit-on-breakeven after 24h if MFE > 1%

🤖 Self-learning status:

• ✅ AUTO-FIX: The model feeds MFE/MAE ratios back into training features. After seeing enough trades where MAE >> MFE, it learns to weight entry conditions that produce BETTER MFE/MAE ratios.
• ⚠️ PARTIAL: Model sees the RSI value as a feature and can learn to avoid entries at RSI > 40, but the FVG strategy's RSI threshold is a HARD FILTER that needs manual adjustment in the strategy code.
• ✅ AUTO-FIX: The model learns from TIME_EXPIRY outcomes that the signal wasn't strong enough for a directional move. Over time it adjusts probability thresholds upward for similar entry conditions.

📋 What else is needed:

• 🔧 MANUAL CHANGE NEEDED: In the FVG strategy config, change rsi_threshold from 50 to 35. Self-learning can't change strategy parameters — only entry probabilities.

Strategy: Support & Resistance Bounce — Price tends to bounce off certain levels repeatedly (support = floor, resistance = ceiling). This strategy buys near sup

Signal reason: Bouncing off support (1.22), distance=2.5%, vol=1.0x

MFE: +4.8% (best price reached) | MAE: -3.2% (worst price reached)

❌ What went wrong:

• Trade expired without hitting TP or SL — no clear directional move materialized. The model was right about no crash, but wrong about an upward move.

🔧 Proposed tweaks:

• Reduce max hold time from 72h to 48h; add exit-on-breakeven after 24h if MFE > 1%

🤖 Self-learning status:

• ✅ AUTO-FIX: The model learns from TIME_EXPIRY outcomes that the signal wasn't strong enough for a directional move. Over time it adjusts probability thresholds upward for similar entry conditions.

📋 What else is needed:

• 📊 Need more data: With only 1 closed trades on this strategy, we need at least 10 to determine if this is a systematic failure or normal variance.

Strategy: Smart Money Fair Value Gap — Tracks where big institutional traders (banks, hedge funds) are likely buying. When price drops into a zone where instit

Signal reason: Bullish FVG fill zone (0.10-0.10), ADX=48, RSI=45

MFE: +0.8% (best price reached) | MAE: -5.2% (worst price reached)

❌ What went wrong:

• RSI was 44 at entry — FVG strategies work best when RSI < 35 (deeply oversold). At 44, price wasn't actually at a discount.
• 2 losses on DOGE-USD — asset was in a sustained downtrend that multiple strategies failed to detect. This means the per-strategy filters weren't enough to catch correlation risk.
• Trade expired without hitting TP or SL — no clear directional move materialized. The model was right about no crash, but wrong about an upward move.

🔧 Proposed tweaks:

• Tighten RSI filter to < 35 for FVG-based strategies
• Add a correlation guard: if 2+ strategies lose on DOGE-USD within 48h, block new entries for 72h
• Reduce max hold time from 72h to 48h; add exit-on-breakeven after 24h if MFE > 1%

🤖 Self-learning status:

• ⚠️ PARTIAL: Model sees the RSI value as a feature and can learn to avoid entries at RSI > 40, but the FVG strategy's RSI threshold is a HARD FILTER that needs manual adjustment in the strategy code.
• ❌ CANNOT SELF-FIX: The self-learning loop operates per-strategy, not cross-strategy. A correlation guard between strategies requires a NEW CODE MODULE that monitors active positions across all strategies. This is a feature gap.
• ✅ AUTO-FIX: The model learns from TIME_EXPIRY outcomes that the signal wasn't strong enough for a directional move. Over time it adjusts probability thresholds upward for similar entry conditions.

📋 What else is needed:

• 🔧 MANUAL CHANGE NEEDED: In the FVG strategy config, change rsi_threshold from 50 to 35. Self-learning can't change strategy parameters — only entry probabilities.
• 🆕 FEATURE NEEDED: Build a CorrelationGuard module in the alpha_engine that tracks open positions per symbol across all strategies and enforces max_positions_per_asset=1.