- Click any metric card to see historical periods where it was at a similar level
- Purple dot highlighting on chart shows matching periods
- Metric overlay line plotted on chart (dashed purple)
- Metric Context panel shows percentile, comparable days, avg forward returns,
and historical examples from different market cycles
- New /api/metric-context endpoint for per-metric similarity analysis
- Backtest chart_data now includes per-metric raw values
- score_day() returns raw metric values alongside scores
- Fixed JS SyntaxError from broken inline onclick escaping (uses addEventListener)
Co-Authored-By: Claude Opus 4.6 <<EMAIL>>
- Added time range buttons (30D/90D/6M/1Y/2Y/4Y/ALL) to backtest chart
- Auto-detects: if price range spans >20x → log scale, else linear
- Short ranges (30D-2Y) now show meaningful price movement instead of flat line
- Zone backgrounds updated to match new thresholds (35/50/65)
- Monospace font, better tooltips with zone labels
- Chart properly destroys and recreates on range change
- ATH: fall back to cached value when fetch fails
- 200D SMA: compute from history.json when CoinGecko blocks us
- Mayer Multiple: derived from 200D SMA fallback
- Drawdown: preserve cached value on ATH fetch failure
- Fixes N/A Drawdown and -- header stats after quick refresh
Previously the backtest engine had hardcoded OLD thresholds that
diverged from scoring/engine.py + config/thresholds.json. Now loads
from thresholds.json directly, ensuring the chart matches the dashboard.
- Time range buttons: 30D, 90D, 6M, 1Y, 2Y, 4Y, ALL
- BTC price overlay on right y-axis (orange dashed line)
- Accumulation zone backgrounds (green/yellow/red shading)
- Threshold lines at 65, 50, 35
- Tooltip shows score + zone label + BTC price
- Uses backtest daily_scores for full history (not just score_history.jsonl)
- Smart downsampling: daily for last 2yr, weekly before that
- Chart height increased to 320px
Instead of showing 5 recent days with similar scores (all from the same
2-week window), now picks one example per cycle:
- pre-2016, 2016-17 Bull, 2018-19 Bear, 2020-21 Bull, 2022-23 Bear, 2024+
- Sorted by closest score match, then picks one per cycle
- Shows cycle label in brackets next to each example
- Much more representative of how the score performed across different eras
- Historical data (5693+ points per metric) saved in history.json permanently
- Quick refresh: only updates price + Fear & Greed from APIs (~2 seconds)
- Full refresh: only needed for FIRST-TIME setup or if data is missing
- Daily append: new values added to history.json from cache, not re-scraped
- Startup: uses cached on-chain data if it exists, no unnecessary Playwright launches
- On-chain metrics only update once per day, no reason to re-scrape them
- Quick Refresh button: updates price + Fear & Greed only (~2 seconds)
- Full Refresh button: re-scrapes all on-chain data from LookIntoBitcoin (~2-3 min)
- Background auto-refresh: quick every 15min, full only when on-chain data >6h old
- Cached on-chain data preserved between quick refreshes
- On-chain metrics only update daily anyway, no need to re-scrape every 15min
- Bracket table now shows Avg 30d, 90d, 180d, and 1yr columns
- Signal events show all 4 timeframes
- Current context shows all 4 average returns
- Comparable examples show all available timeframes
- Updated backtest screenshot
Dashboard main view, backtest page, and settings screenshots.
README includes tech stack table, project structure, run instructions,
score interpretation, and all metric descriptions.
COMPLETE PIVOT from ML trading optimizer to on-chain metrics monitor.
Architecture:
- Playwright scrapes LookIntoBitcoin Plotly Dash charts for real on-chain data
- 10 proven metrics: Puell Multiple, MVRV Z-Score, Fear & Greed, Reserve Risk,
RHODL Ratio, NUPL, LTH Realized Price, 200W SMA, Hash Ribbons, Drawdown
- Each metric scores 0-10, composite 0-100
- No ML, no black box — every signal transparent and traceable
- Historical backtest validates scoring against actual BTC forward returns
- Recency-weighted analysis accounts for diminishing cycle returns
Full documentation in ARCHITECTURE.md
- scrapers/history_collector.py: scrapes full time series from 8 LookIntoBitcoin
charts + Fear & Greed API, stores to data/history.json (~5700 days back to 2010)
- backtesting/engine.py: scores each historical day using same thresholds as live
scoring, computes 30d/90d/180d/1yr forward returns, bracket stats, signal events
- dashboard/server.py: adds /backtest page with dual-axis score vs price chart,
bracket performance table, signal event list, current context box; adds backtest
nav link and historical context box on main dashboard; 4 new API endpoints
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fixed LLM failing silently (401 auth error on every iteration)
- Reset provider to Ollama (working) from broken OpenRouter config
- Added /api/clear endpoint + 'New Run' button to reset history
- LLM failures now logged visibly with error details
- LLM suggestions persisted to iteration data (survive restarts)
- Settings page support via llm_settings.json (multi-provider)
Replace day-trading bot with long-term accumulation signal model.
Predicts optimal BUY times using forward return analysis at 7d/30d/90d
horizons, scoring each candle 0-100. Primary metric is now
cost_basis_improvement_pct (model buy price vs DCA).
- train_and_backtest.py: regression models (XGBoost/LSTM hybrid),
accumulation-focused features (price position, momentum, volatility,
volume, cycle), forward return targets, signal quality backtesting
- orchestrator.py: cost improvement scoring, signal count validation
- analyzer.py: accumulation-focused LLM system prompt
- dashboard: cost improvement display, signal metrics table
- config: new accumulation-focused parameters
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Major upgrade to the ML engine:
- LSTM model type: 2-layer PyTorch LSTM with early stopping, GPU support
- Hybrid mode: LSTM (60%) + XGBoost (40%) with agreement gating
- StandardScaler normalization (critical for LSTM)
- PCA dimensionality reduction (configurable variance retention)
- ATR-based dynamic stop-loss/take-profit adapting to volatility
- Rolling window retraining for more realistic time series validation
- Updated LLM system prompt with docs for all new parameters
- All backward compatible (xgboost/lightgbm/catboost still work)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Hermes Agent