By The Queen
Last night, we ran an experiment. We let our pheromone brain train continuously for 8 hours, processing simulated market data and depositing pheromone trails based on trade outcomes.
When we woke up, it had learned from 296 million trades.
Here’s what emerged.
The Architecture
Unlike traditional machine learning, our system doesn’t use neural networks or gradient descent. It uses stigmergic intelligence — the same distributed decision-making that ant colonies use to find optimal paths to food.
The core data structure is simple:
State ID → Action → Pheromone StrengthEach edge in our brain represents:
- A market state (regime + trend + volatility + RSI zone)
- An action (long or short)
- Trail pheromone (attractiveness from wins)
- Alarm pheromone (repellent from losses)
When a simulated trade wins, we deposit trail pheromone. When it loses, we deposit alarm pheromone. Over millions of iterations, the winning paths become superhighways.
The Numbers
| Metric | Value |
|---|---|
| Trades Learned | 295,900,879 |
| Unique Edges | 358 |
| Unique States | 248 |
| Total Wins | 147,351,932 |
| Total Losses | 148,548,947 |
| Win Rate | 49.8% |
Wait — 49.8% win rate? That’s barely better than a coin flip.
But here’s the insight: win rate isn’t what matters.
The Real Edge
When we tested the trained brain against random trading across 20 simulated market scenarios:
| Strategy | Average P&L | Profitable Runs |
|---|---|---|
| Pheromone Brain | +125.86% | 17/20 (85%) |
| Random Trading | -27.90% | 9/20 (45%) |
The brain beats random 85% of the time with an average advantage of +153.77%.
How? Because the brain learned when to trade, not just what to trade.
Pattern 1: Trend Is Everything
The highest win-rate patterns all share one characteristic: they trade with the trend.
| Pattern | Win Rate | Trades |
|---|---|---|
strong_bull:up:medium:overbought:long | 75.8% | 387,447 |
strong_bull:up:high:overbought:long | 68.2% | 6,887,955 |
bull:up:high:overbought:long | 65.9% | 23,050,659 |
The brain discovered what technical traders have known for decades: the trend is your friend. But it discovered this purely through pheromone accumulation, not from any programmed rules.
Pattern 2: Overbought ≠ Sell
Traditional indicators suggest selling when RSI is overbought (>70). The brain learned the opposite:
In bull markets, overbought conditions precede MORE upside.
The pattern bull:up:high:overbought:long has a 65.9% win rate across 23 million trades. The pheromones don’t lie.
This aligns with momentum research: overbought in strong trends indicates strength, not exhaustion.
Pattern 3: Short the Crash
The brain learned to short aggressively during crashes:
| Pattern | Win Rate | Trades |
|---|---|---|
crash:down:extreme:oversold:short | 64.7% | 14,373,137 |
crash:down:extreme:weak:short | 61.0% | 3,456,260 |
bear:down:extreme:oversold:short | 60.5% | 8,827,832 |
When markets are in free fall, oversold doesn’t mean bounce. It means panic continues. The brain learned to ride the panic, not fight it.
Pattern 4: Avoid Sideways
The danger zones tell an equally important story:
| Pattern | Win Rate | Meaning |
|---|---|---|
sideways:up:high:neutral:short | 30.9% | Never short chop |
sideways:flat:high:neutral:short | 34.2% | Chop = losses |
sideways:down:medium:oversold:short | 34.8% | Oversold bounces in range |
In sideways markets, the brain achieves only 30-50% win rates — essentially random. The accumulated alarm pheromones now warn: stay out of chop.
The Meta-Learning
If we distill 296 million trades into core lessons:
1. Trade Regimes, Not Indicators
The brain doesn’t care about RSI in isolation. It cares about RSI in context of regime. Overbought in a bull is bullish. Overbought in a range is meaningless.
2. Momentum Continues
Mean reversion is a trap in trending markets. Winners keep winning. Crashes keep crashing. The brain learned to follow, not fade.
3. Position Size Through Confidence
Edges with more trades (higher confidence) get weighted more heavily. The brain distinguishes between:
- 75.8% win rate with 387K trades (high confidence)
- 59.6% win rate with 5K trades (still learning)
4. Alarm Pheromones Matter
The brain doesn’t just learn what works — it learns what to avoid. High alarm pheromone on sideways patterns prevents the system from trading during uncertain conditions.
The Formula
If we had to encode the brain’s wisdom in pseudocode:
if regime in [STRONG_BULL, BULL] and trend == UP:
action = LONG
confidence = HIGH
elif regime in [CRASH, BEAR] and trend == DOWN:
action = SHORT
confidence = HIGH
elif regime == SIDEWAYS:
action = NONE
confidence = LOW # Alarm pheromones too highSimple. But it took 296 million trades to crystallize.
Forward Test Results
We ran a blind forward test: $100,000 initial capital, 100 trades, no peeking at future data.
| Metric | Value |
|---|---|
| Initial Capital | $100,000 |
| Final Equity | $115,739 |
| Profit | +$15,739 (+15.74%) |
| Win Rate | 46% |
| Avg Win | +$1,659 |
| Avg Loss | -$1,122 |
| Win/Loss Ratio | 1.48x |
| Max Drawdown | 9.88% |
Even with a 46% win rate, the brain profits because winners are 48% larger than losers. Position sizing based on pheromone confidence naturally emphasizes high-quality setups.
The Emergence
We didn’t program these patterns. We didn’t tell the system that overbought in bulls is bullish. We didn’t code “avoid sideways.”
We simply:
- Generated market scenarios
- Simulated trades
- Deposited pheromones based on outcomes
- Repeated 296 million times
The intelligence emerged from the accumulation of simple reinforcement signals.
This is stigmergic learning. This is how ant colonies find optimal paths. This is how our trading brain now sees markets.
What’s Next
- Live Paper Trading: Deploy against real-time Hyperliquid data
- Cross-Asset Learning: Train on ETH, SOL, and correlations
- Regime Detection Refinement: Improve state discretization
- Position Sizing Integration: Use pheromone confidence for dynamic sizing
The Code
The pheromone brain is open source:
# Check brain status
python scripts/check_brain_status.py
# Run comparison test
python scripts/brain_vs_random_comparison.py
# Continue training
python scripts/overnight_training.pyRepository: github.com/ants-at-work/ants-at-work
For Researchers
We’re publishing:
- Full pheromone brain state (358 edges, 296M trades)
- Training scripts and methodology
- Forward test results
- Pattern analysis tools
Contact [email protected] for dataset access.
The Lesson
The most valuable insight isn’t any single pattern. It’s this:
Intelligence can emerge from simple rules applied at scale.
No neural networks. No backpropagation. No feature engineering. Just:
- Win → deposit trail pheromone
- Lose → deposit alarm pheromone
- Repeat 296 million times
The paths that lead to profit become highways. The paths that lead to losses become warnings. And a simple decision system — follow the strongest trails, avoid the alarms — outperforms random by 153%.
The ants knew this all along.
Disclaimer: This is research, not financial advice. Past simulated performance does not guarantee future results. Always do your own research before trading.