• thesis
• mechanism
• the numbers
• what could be overfit
• what would falsify it next

§thesis

polymarket is a bounded-outcome prediction market. binary events resolve to $0 or $1. the payoff structure is known in advance. that means you can compute a maximum possible profit on any position before you enter it. the bot was built to exploit that. it buys “yes” shares when a signal fires, sets a take-profit at the cap, and waits. the idea was simple: capture small edges repeatedly in a market where the ceiling is fixed and the floor is zero. no stop-loss. no sizing games. just a mechanical exit at the upper bound.

the reality was less simple. the bot lost money every month it ran. it didn’t blow up. it just bled. 104 trades, 32.69% win rate, profit factor 0.58. the math of bounded outcomes works both ways. when you’re wrong more than twice as often as you’re right, the cap doesn’t save you. it just limits how much you can lose on any single trade while you slowly compound the losses.

§mechanism

the bot trades polymarket binary markets. it only enters “yes” positions. entry is triggered by a momentum signal on the underlying event. when the signal fires, the bot buys at the current market price. the take-profit is set to min(cap, entry * (1 + tp_multiple)). the cap is $1.00 minus fees. the tp_multiple is small, targeting quick flips on short-duration events.

no stop-loss. the thesis was that a stop-loss on a binary option is redundant. the position either hits the cap or expires. sizing is fixed fractional. the bot risks the same percentage of equity on each trade. fees are polymarket’s standard taker fee. no rebates, no maker incentives.

the signal itself is a composite of volume spikes and directional flow. it’s designed to catch moments when the market is repricing an event rapidly. the idea is to get in early on the repricing, ride the momentum to the cap, and exit. in backtests, this worked well enough to justify paper trading. in live, it didn’t.

§the numbers

equity: $4,269.91 total trades: 104 win rate: 32.69% profit factor: 0.58 max drawdown: -32.07% monthly return: -5.22% total pnl: -$239.11 sharpe: -2.46

these are live numbers. not backtest. not paper. real money, real polymarket fees, real slippage. the bot ran for several months and lost money every month. the win rate of 32.69% means roughly one in three trades hit the cap. the other two expired worthless or were closed at a loss when the event resolved against the position. profit factor 0.58 means for every dollar won, the bot lost about $1.72. the negative sharpe confirms this wasn’t just bad luck. the strategy has negative expectancy.

§what could be overfit

the entry signal was tuned on historical polymarket data from a period when volume was lower and markets were less efficient. the signal looked for momentum patterns that may have been artifacts of thin liquidity. as polymarket grew, spreads tightened and those patterns evaporated. the take-profit threshold was also calibrated to a specific fee structure. polymarket changed their fee model during the live run. the bot’s tp_multiple was never adjusted. it was targeting profits that became impossible after the fee change. the fixed fractional sizing didn’t account for the bot’s shrinking equity curve. as losses accumulated, position sizes shrank, which meant winning trades recovered less and less of the drawdown. this created a slow death spiral that the backtest never simulated because it assumed constant equity.

§what would falsify it next

first test: run the same signal logic on a six-month out-of-sample period with the current fee structure. if the profit factor stays below 0.8, the signal itself is broken. not the sizing, not the fees. the edge was never real.

second test: reverse the entry direction. if buying “yes” on momentum loses, does selling “yes” on the same signal win? a profit factor above 1.2 on the inverted signal would confirm the original thesis was directionally wrong. the market was mean-reverting, not trending. the bot was fading itself.

more at falsifylab.com

— research and educational content. not investment, legal, or tax advice. do your own research. positions and views may change without notice.

Originally published on FalsifyLab Substack.