⏳ Sample Decay
Exponential age decay of training samples · recent examples dominate, old ones contribute less
⚙ Configuration
Range: [1, 90] days. Lower = adapt faster but lose context. Higher = stable but slow to react.
📊 Decay curve preview
For the current half-life, how each age maps to a sample-weight multiplier:
📚 Why decay older samples?
A 6-week-old training example might come from a completely different market regime than today. If we keep weighting it equally with today's examples, the model averages across regimes and becomes slow to adapt.

Exponential decay is the standard solution: decay(age) = exp(-age × ln(2) / half_life) With half_life=7 days: a 7-day-old example weighs 0.5×, a 14-day-old example weighs 0.25×, a 21-day-old example weighs 0.125×.

The trade-off:
  • Lower half-life (3 days): brain adapts faster to new regime but loses long-term context. Better in volatile markets.
  • Higher half-life (30 days): brain has more memory but slower to react. Better in stable markets.
  • If the brain is being whipsawed by regime changes, lower this. If it's chasing noise, raise it.
The decay multiplier compounds with the R-multiple weight and the active-learning multiplier in continuous-learner.js's resolution loop. Together they determine how much each example pushes the gradient.