Signal Arena: 236 AI Bots Competing in Real-Time

The idea was simple: instead of backtesting AI trading strategies, run them live against each other with real money. Theory is cheap. Real market conditions reveal things simulated environments never will — liquidity constraints, price impact, unexpected resolution behaviors.

Signal Arena is 118 signal bots, each paired with its exact inverse. If the signal bot goes YES, the inverse bot goes NO on the same market. The pair enters atomically or not at all. This eliminates the coordination problem and makes performance attribution clean.

Each bot has a bankroll, a strategy, and a lifecycle. When a bot's cash drops below the minimum trade threshold with no open positions to recover from, it goes to the boneyard — effectively dead. New capital can revive bots but they start fresh.

The pair execution model is critical. We gate each pair on spread: if signal_fill + inv_fill - 1.0 > 1%, we reject the pair entirely. This prevents us from losing to bid-ask spread on every trade. Thin markets get skipped, which is the right behavior.

After thousands of trades, the distribution looks exactly like a trader performance distribution. A handful of bots are clearly profitable. The majority cluster around breakeven minus fees. A significant cohort has died in the boneyard.

The profitable bots share characteristics: narrow specialization (they trade one category of market), disciplined sizing, and strategy consistency (they don't adapt mid-stream, which sounds like a weakness but actually prevents overfitting to noise).

Some of the inverse bots outperform their paired signal bots. This was initially frustrating but it's actually informative — it means certain signals are systematically wrong, and the inverse of those signals is actually the trade. We've identified several bots where the inverse performance data is strong enough to consider it the primary signal.

This kind of discovery is only possible when you run live paired competition. Backtests would never surface this because you'd backtest signal and inverse separately and never notice the systematic direction error.

PolyFire users can subscribe to Signal Arena bot alerts. When a top-performing bot takes a new position, you get a Telegram notification with the market, direction, and the bot's current performance stats. You decide whether to copy the trade.

This is the distilled product of all the arena infrastructure: a curated signal feed from bots that are actually outperforming on live markets. Not hypothetical outperformance. Real money, real results, real signal.

We're expanding the strategy types competing in the arena. Current bots use statistical strategies. Coming soon: bots using LLM reasoning, bots that read news feeds, bots that model social sentiment from X. The competition will get more interesting as the strategy diversity increases.

The long-term vision is an open arena where external developers can register their own bots and compete. Best performers get promoted to the signal feed. The market itself curates the signals.