Train the brain on REAL past market data · pulled from Stooq · honest features and outcomes · --
What this does: Fetches multi-year daily OHLC bars for every symbol from Stooq (free, no key). For each historical day computes 22 features from REAL bars (RSI from real closes, ATR from real ranges, MA distances, etc.). Labels: did next-day return exceed 0.3×ATR%? Splits data 80/20 train/test (last 14 days = test). Trains 3 epochs of SGD. Reports HELD-OUT accuracy and compares to a majority-class baseline.
Why honest reporting matters: The original auto-seed generated 200 random rows and the model 'looked smart' from minute one — but its 'training loss' was meaningless because the labels were also random. This version reports test accuracy on a held-out period the model never saw. If accuracy doesn't beat baseline, there's no edge and the dashboard will say so.
Cost: 22 HTTP requests to Stooq. ~6 seconds. Run weekly to refresh on new bars.
Symbols available
22
Bars fetched
0
Training examples
0
Test accuracy
—
~5 seconds. Tags every row dataSource='live' so the live-only trainer accepts them.
📊 Training loss curve
📋 Log
🔬 What the model learns
For each historical day per symbol, we compute:
• RSI(14) from real close history
• ATR% from real daily ranges
• Distance from 20/50 MAs from rolling averages
• Price-relative-to-prevClose %
• Day range / body / wick as % of ATR
The label is the simplest possible signal: did the close 5 trading days later exceed today's close? Yes = label 1 (win), No = label 0 (loss). The brain learns which feature combinations predict that.
After training, the brain has real ground-truth knowledge of "what RSI + MA distance combinations preceded next-week winners over the last 90 days" — actual signal from actual market behavior.