Learned Weights
How the scanner teaches itself which indicators to trust.
Learned Weights are the scoring factors of the grid and futures scanners whose values are adjusted automatically by real trading results, instead of being fixed by hand. Each factor starts at an equal weight (0.25); after every closed position, factors that predicted successful trades gain weight and factors that mislead lose weight.
Why weights at all?
Every scan ranks coins with a composite score. For the grid scanner the factors are oscillation, ADX, drift, and 24h width; for the futures scanner they are RSI, momentum, Z-score, and volatility. The score is a weighted sum:
score = w₁·factor₁ + w₂·factor₂ + w₃·factor₃ + w₄·factor₄
Some factors are better predictors than others — but nobody knows in advance which. Learned weights let the results decide.
The feedback loop
- Record the setup: when a coin is scored and traded, the factor values behind the signal are stored.
- Watch the outcome: did the grid earn profit? Did the futures position hit its take-profit or stop-loss?
- Attribute rewards: each factor earns a reward based on how well its signal correlated with the actual result.
- Update the weights: factors with positive rewards gain weight; factors with negative rewards lose weight. The weights are renormalized so they always sum to 1.
w_i(new) = w_i(old) × e^(η × reward_i)
Each closed position is one training cycle. Over time, the weights encode the strategy's real empirical edge.
Worked example
| Factor | Reward | Old weight | New weight |
|---|---|---|---|
| Z-score | +0.6 | 0.25 | 0.28 |
| Momentum | +0.2 | 0.25 | 0.25 |
| Volatility | −0.4 | 0.25 | 0.21 |
| RSI | 0.0 | 0.25 | 0.26 |
Z-score correlated with winners, so it gains influence; volatility misbehaved, so it shrinks. The system has learned — without anyone hard-coding it.
Design choices in Blind Monkeys
- Slow adaptation: a small learning rate (η) means weights move a little per trade — stability over speed.
- Correlation risk: factors that rise and fall together can be penalized or rewarded as a group, even if only one was truly responsible.
- Reward definition matters: for grids, "good" means oscillation density, contained drawdown, and range persistence — not just a raw take-profit hit.
- Visible history: the current learned weights are displayed on the scanner pages and update after each closed position.