Enhancement #6 - SL Distance Validation (Data Collection Phase): - Added slDistanceAtEntry field to StopHunt schema - Calculates distance from revenge entry to stop zone (LONG vs SHORT logic) - Logs distance in dollars + × ATR multiplier - Purpose: Collect 20+ revenge trade samples for optimal multiplier analysis - Created comprehensive analysis guide with SQL queries - Decision deferred until empirical data collected Enhancement #1 - ADX Confirmation (Implementation Plan): - Documented complete 1-minute TradingView alert strategy - Storage analysis: 19.44 MB/month for 3 symbols (negligible) - Two-phase approach: Cache-only MVP → Optional DB persistence - Provided TradingView Pine Script (ready to use) - Cost breakdown: Pro subscription $49.95/month required - Benefits: Real-time ADX, pattern recognition, ML features - Implementation checklist with validation phases Files Changed: - prisma/schema.prisma: +1 field (slDistanceAtEntry) - lib/trading/stop-hunt-tracker.ts: +10 lines (distance calculation + logging) - docs/1MIN_MARKET_DATA_IMPLEMENTATION.md: NEW (comprehensive plan) - docs/ENHANCEMENT_6_ANALYSIS_GUIDE.md: NEW (SQL queries + decision matrix) Status: Enhancement #4 and #10 deployed (previous commit) Enhancement #6 data collection enabled (this commit) Awaiting 20+ revenge trades for Enhancement #6 decision
9.9 KiB
Enhancement #6 - SL Distance Analysis Guide
Status: DATA COLLECTION ENABLED (Nov 27, 2025)
Purpose: Gather revenge trade data to determine optimal SL distance multiplier
What We're Tracking Now
Database Fields Added
-- In StopHunt table:
slDistanceAtEntry Float? -- Distance from entry to stop zone (absolute value)
revengeOutcome String? -- "TP1", "TP2", "SL", "TRAILING_SL"
revengePnL Float? -- Actual P&L from revenge trade
revengeFailedReason String? -- "stopped_again" if re-stopped
originalATR Float? -- ATR at original stop-out
What Gets Calculated
// When revenge trade executes:
const slDistance = direction === 'long'
? currentPrice - stopHuntPrice // LONG: Room below entry
: stopHuntPrice - currentPrice // SHORT: Room above entry
// Stored: Math.abs(slDistance)
// Logged: Distance in $ and in ATR multiples
// Example: "$1.48 distance (2.47× ATR)"
Analysis Queries (After 20+ Revenge Trades)
Query 1: SL Distance vs Outcome
-- Do tighter entries get re-stopped more often?
SELECT
CASE
WHEN "slDistanceAtEntry" / "originalATR" < 1.0 THEN '<1× ATR (Very Tight)'
WHEN "slDistanceAtEntry" / "originalATR" < 1.5 THEN '1-1.5× ATR (Tight)'
WHEN "slDistanceAtEntry" / "originalATR" < 2.0 THEN '1.5-2× ATR (Moderate)'
WHEN "slDistanceAtEntry" / "originalATR" < 2.5 THEN '2-2.5× ATR (Safe)'
ELSE '2.5×+ ATR (Very Safe)'
END as distance_tier,
COUNT(*) as total_trades,
-- Win rate
ROUND(100.0 * SUM(CASE
WHEN "revengeOutcome" IN ('TP1', 'TP2', 'TRAILING_SL') THEN 1
ELSE 0
END) / COUNT(*), 1) as win_rate,
-- Re-stopped rate
ROUND(100.0 * SUM(CASE
WHEN "revengeFailedReason" = 'stopped_again' THEN 1
ELSE 0
END) / COUNT(*), 1) as restopped_rate,
-- Average P&L
ROUND(AVG("revengePnL"), 2) as avg_pnl,
-- Total P&L
ROUND(SUM("revengePnL"), 2) as total_pnl
FROM "StopHunt"
WHERE "revengeExecuted" = true
AND "slDistanceAtEntry" IS NOT NULL
AND "originalATR" IS NOT NULL
GROUP BY distance_tier
ORDER BY MIN("slDistanceAtEntry" / "originalATR");
Expected Output:
distance_tier | total | win_rate | restopped_rate | avg_pnl | total_pnl
-----------------------+-------+----------+----------------+---------+-----------
<1× ATR (Very Tight) | 5 | 40.0 | 60.0 | -25.50 | -127.50
1-1.5× ATR (Tight) | 8 | 62.5 | 25.0 | 15.25 | 122.00
1.5-2× ATR (Moderate) | 12 | 75.0 | 16.7 | 42.30 | 507.60
2-2.5× ATR (Safe) | 7 | 85.7 | 14.3 | 68.45 | 479.15
2.5×+ ATR (Very Safe) | 3 | 100.0 | 0.0 | 92.30 | 276.90
Decision Logic:
- If <1× ATR has high re-stop rate (>50%): Filter needed
- If 1.5-2× ATR has best risk/reward: Use 1.5× multiplier
- If 2-2.5× ATR has highest win rate: Use 2.0× multiplier
- If Very Safe (2.5×+) rarely happens: Lower multiplier to catch more
Query 2: Missed Opportunities
-- How many revenge opportunities didn't execute because reversal wasn't deep enough?
SELECT
COUNT(*) as total_stop_hunts,
-- How many reversed but didn't enter
SUM(CASE
WHEN "revengeExecuted" = false
AND "revengeWindowExpired" = true
AND ("lowestPriceAfterStop" < "originalEntryPrice" -- LONG reversed
OR "highestPriceAfterStop" > "originalEntryPrice") -- SHORT reversed
THEN 1 ELSE 0
END) as missed_reversals,
-- Average distance of missed reversals
ROUND(AVG(CASE
WHEN "revengeExecuted" = false
AND "revengeWindowExpired" = true
THEN ABS("lowestPriceAfterStop" - "stopHuntPrice")
END), 2) as avg_missed_distance
FROM "StopHunt"
WHERE "originalATR" IS NOT NULL;
Tells us: If we filter too strictly (e.g., 3× ATR), how many opportunities do we lose?
Query 3: Time to Re-Stop
-- How quickly do re-stopped revenge trades fail?
SELECT
CASE
WHEN EXTRACT(EPOCH FROM ("Trade"."exitTime" - "StopHunt"."revengeTime")) < 300 THEN '<5min (Instant)'
WHEN EXTRACT(EPOCH FROM ("Trade"."exitTime" - "StopHunt"."revengeTime")) < 900 THEN '5-15min (Fast)'
WHEN EXTRACT(EPOCH FROM ("Trade"."exitTime" - "StopHunt"."revengeTime")) < 1800 THEN '15-30min (Moderate)'
ELSE '30min+ (Slow)'
END as time_to_restop,
COUNT(*) as count,
ROUND(AVG("StopHunt"."slDistanceAtEntry" / "StopHunt"."originalATR"), 2) as avg_atr_multiple
FROM "StopHunt"
INNER JOIN "Trade" ON "StopHunt"."revengeTradeId" = "Trade"."id"
WHERE "StopHunt"."revengeFailedReason" = 'stopped_again'
GROUP BY time_to_restop
ORDER BY MIN(EXTRACT(EPOCH FROM ("Trade"."exitTime" - "StopHunt"."revengeTime")));
Insight: If most re-stops happen <5min, they're wicks. Wider SL distance helps.
Query 4: Direction-Specific Analysis
-- Do LONGs vs SHORTs need different SL distances?
SELECT
"direction",
COUNT(*) as total,
ROUND(AVG("slDistanceAtEntry" / "originalATR"), 2) as avg_atr_multiple,
ROUND(100.0 * SUM(CASE
WHEN "revengeFailedReason" = 'stopped_again' THEN 1 ELSE 0
END) / COUNT(*), 1) as restopped_rate,
ROUND(AVG("revengePnL"), 2) as avg_pnl
FROM "StopHunt"
WHERE "revengeExecuted" = true
AND "slDistanceAtEntry" IS NOT NULL
GROUP BY "direction";
Possible outcome: SHORTs need wider distance (more volatile reversals)
Decision Matrix (After 20+ Trades)
Scenario A: Tight Entries Work
Data shows: 1-1.5× ATR has 70%+ win rate, low re-stop rate Decision: Use 1.5× ATR multiplier (or no filter at all) Trade-off: More revenge opportunities, slightly higher re-stop risk
Scenario B: Moderate Distance Optimal
Data shows: 1.5-2× ATR has best risk/reward Decision: Use 1.75× or 2.0× ATR multiplier Trade-off: Balanced approach (recommended starting point)
Scenario C: Wide Distance Required
Data shows: <2× ATR has >40% re-stop rate Decision: Use 2.5× or 3.0× ATR multiplier Trade-off: Fewer opportunities, but much higher win rate
Scenario D: Direction-Specific
Data shows: LONGs work at 1.5×, SHORTs need 2.5× Decision: Implement separate multipliers per direction Trade-off: More complex but optimized
Implementation Plan (After Data Collection)
Step 1: Review Data (After 20 Revenge Trades)
# Run Query 1 in database
docker exec trading-bot-postgres psql -U postgres -d trading_bot_v4 -c "
[Query 1 from above]
"
# Save results to CSV
# Analyze in spreadsheet or share with me
Step 2: Calculate Optimal Multiplier
# Simple calculation:
optimal_multiplier = (
sum(distance * pnl for each tier) /
sum(pnl for each tier)
)
# Weight by win rate:
optimal_multiplier_weighted = (
sum(distance * win_rate * count for each tier) /
sum(win_rate * count for each tier)
)
Step 3: Implement Filter
// In stop-hunt-tracker.ts shouldExecuteRevenge()
const slDistance = stopHunt.direction === 'long'
? currentPrice - stopHunt.stopHuntPrice
: stopHunt.stopHuntPrice - currentPrice
const minSafeDistance = stopHunt.originalATR * 2.0 // Use data-driven value
if (Math.abs(slDistance) < minSafeDistance) {
console.log(`⚠️ SL distance too tight: ${Math.abs(slDistance).toFixed(2)} < ${minSafeDistance.toFixed(2)}`)
return false
}
Step 4: A/B Test (Optional)
// Randomly assign filter on/off
const useFilter = Math.random() > 0.5
if (useFilter && Math.abs(slDistance) < minSafeDistance) {
await prisma.stopHunt.update({
where: { id: stopHunt.id },
data: {
notes: 'Would have been filtered by SL distance check',
filterTestGroup: 'A'
}
})
return false
}
// Compare groups after 40 trades (20 each)
Monitoring Dashboard
Key Metrics to Track
- Re-Stop Rate: % of revenge trades that hit SL immediately
- Average SL Distance: Median ATR multiple at entry
- Win Rate by Distance: Scatterplot (x=distance, y=outcome)
- Missed Opportunities: Reversals that didn't execute
Alert Thresholds
- High re-stop rate: >30% revenge trades fail = tighten filter
- Low execution rate: <50% stop hunts execute = loosen filter
- Negative total P&L: Sum of revenge P&L < 0 = pause system
Timeline
Week 1-2: Data collection (first 5-10 revenge trades) Week 3: Preliminary analysis (spot obvious patterns) Week 4: Full analysis (20+ trades, statistical significance) Week 5: Implement optimal filter based on data Week 6+: Monitor improvement, adjust if needed
Current Status: ✅ Data collection enabled, awaiting first revenge trades
Example Real-World Analysis
Hypothetical Results After 25 Revenge Trades
Distance Tier | Trades | Win Rate | Avg P&L | Total P&L
---------------------|--------|----------|---------|----------
<1× ATR | 3 | 33% | -$45 | -$135
1-1.5× ATR | 7 | 57% | $12 | $84
1.5-2× ATR | 9 | 78% | $58 | $522
2-2.5× ATR | 4 | 75% | $62 | $248
2.5×+ ATR | 2 | 100% | $89 | $178
Conclusion from data:
- <1× ATR: High failure rate, negative P&L → FILTER OUT
- 1-1.5× ATR: Break-even, mediocre win rate → QUESTIONABLE
- 1.5-2× ATR: SWEET SPOT - 9 trades, 78% WR, +$522
- 2-2.5× ATR: Good but fewer opportunities
- 2.5×+ ATR: Perfect WR but rarely happens
Optimal setting: 1.5× ATR minimum distance Expected impact: Block 10 bad trades, keep 15 good trades = +$606 net improvement
Questions for Future AI Agents
- What's the optimal multiplier? → Check Query 1 results
- Are we missing too many opportunities? → Check Query 2 results
- Do re-stops happen instantly or slowly? → Check Query 3 results
- Do LONGs vs SHORTs need different filters? → Check Query 4 results
- Should we implement the filter? → If re-stop rate >30%, YES
Remember: Data > theory. Don't implement filter until data proves it helps.