📊 Version-Specific Statistics Accuracy — Jul 16
Improved statistical aggregation and display for each ML version (v3.2 / v3.3 / v3.4). The detail page now shows scores and grades corresponding to the selected version, and signals from June 20 onwards are now correctly included.
🎨 UI Consistency & Usability — Jul 16
Unrelated top selectors are now automatically hidden during heatmap display. Model version labels have been unified from internal identifiers to user-friendly format (v3.2 / v3.3 / v3.4). Version badges in the detail view now reflect the selected version.
🧠 Summary "ML-Filtered Win Rate" Unified — Jul 16
The summary card's "ML-filtered win rate" now uses the same criteria as hourly, weekday, and daily charts. This corrects the number of qualifying signals for v3.4 to statistically meaningful values.
🔗 Detail Page Filter Preservation — Jul 16
After navigating to the detail page from hourly, weekday, daily, or LLM tabs, changing filters (period, sort, model version, etc.) no longer removes the original filter (e.g., 5 AM, Monday).
⚠️ Notice: v3.3 Model 15min/30min Prediction Adjustment — Jul 16
We identified that the confidence score distribution for v3.3 model 15-minute and 30-minute predictions was concentrated in a narrower range than expected, causing grade classifications to skew toward C. Prediction direction (Bullish / Bearish) has been functioning correctly with no impact on win rates. We are currently working on model adjustments and will provide an update once the fix is complete. Historical data shown in reports is preserved as-is for transparency and reliability.