ga_trader / genome.py + features.py

What one trader's brain actually looks like

Every trader in the population is one tiny MLP plus a handful of evolved scalar genes riding alongside the weights. No backprop — crossover and mutation shape everything below. Two architectures actually run in this project; toggle between them.

Network

tanh(W₁x + b₁) → softmax(W₂h + b₂)

33 hand-engineered features, 200-bar lookback 12 hidden units, tanh 3 actions, softmax

How a softmax probability becomes a trade

engine.py — Market.step()

Forward pass

Batched across all 2,000 traders at once, every tick. Same features, 2,000 different evolved weight sets.

Confidence gate

Acts only if max(prob) ≥ conf and its margin over "flat" clears margin — both evolved per trader.

RSI gate

If rsi_gate > 0.55, blocks buying overbought / shorting oversold — an optional evolved guard.

Position sizing

Notional = cash equity × 2× leverage × size gene, frozen at entry.

Exit

Whichever comes first: stop_frac adverse move, max_hold_sec elapsed, or a confident reversal signal.

The genes — evolved, not trained

genome.py — random_genome() / mutate()

conf

0.34 – 0.62

Minimum softmax confidence required before acting on any signal at all.

margin

0.04 – 0.22

How far the top action's probability must clear "flat" — decisiveness, not just confidence.

stop_frac

0.004 – 0.06

Adverse price move (as a fraction) that force-closes the position early.

max_hold_sec

45s – 5min *

Hard ceiling on how long a position can stay open, whatever else happens.

size

0.35 – 1.0

Fraction of allowed notional (cash × leverage) committed when a position opens.

rsi_gate

0.0 – 1.0

Threshold controlling whether the optional overbought/oversold guard is active.

* max_hold_sec's ceiling is itself evolvable per-experiment — 5 minutes live, up to 8 hours in the day-trade experiments. Every value above is the true range at random init; mutation can drift it within its bounds thereafter.
Population: 2,000 traders, one forward pass, fully vectorized — forward_batch() in genome.py. Evolution is cull → breed (crossover + mutation) → repeat, never gradient descent.