🧱 Per-Symbol Meta-Stacker
Each symbol learns its own ensemble blend · NVDA's optimal mix differs from SPY's
🏆 Symbol training leaderboard
Each symbol needs ≥30 resolutions before its own per-symbol weights kick in. Below that threshold, the predictor falls back to the global MetaStacker.
Weights shown are softmax-normalized — they represent the relative trust the per-symbol blend gives to each base learner.
📚 Why per-symbol blends
The global MetaStacker (meta-stacker.html) learns ONE optimal blend across all symbols. But different symbols have different dynamics:
  • SPY — large-cap index, smooth dynamics → multi-horizon ensemble often dominates
  • NVDA / TSLA — high-vol momentum names → bootstrap divergence carries more info
  • Crypto (BTC/ETH) — 24/7 markets → k-NN of similar past states may be most predictive
Per-symbol meta-stackers learn each symbol's optimal mix independently. After 30+ resolutions per symbol, the brain switches that symbol's predictions to its own learned blend. Below that threshold, the global blend is used.

Storage: ~28 bytes per symbol (6 weights + bias). 24 symbols = ~700 bytes total. Trivial overhead for genuinely better predictions.