The Trader Behind the Bot

The most useful thing an AI trading bot can do is fail. Not crash — fail precisely, in ways you can diagnose.

Today PolyEdge scored 65/100 on three consecutive setups and declined to trade all three. The 80-point MODERATE threshold held firm. The bot sat. I traded all three manually and went 3 for 3. Now I'm shipping two PRs that will close the gaps the bot just revealed.

This is how you improve a model that has real money behind it.

What the Bot Saw vs. What I Saw

PolyEdge's scoring architecture pulls from nine factors — InDecision Framework as the backbone, fed by live Coinbase price data, MACD readings across timeframes, volume, Chainlink oracle price references, and window timing context. When the 9-factor composite lands below 80, the bot skips. That threshold exists because anything below it historically produces negative expected value after Polymarket's fee structure.

Today's BTC daily candle was a statement. Feb 25 opened with a hammer reversal, followed by a full engulfing bull candle. BTC ran +6.67% on the day, closing at $68,336. The daily structure was unambiguous.

The bot saw the daily momentum. score_momentum uses multi-period rate-of-change across timeframes — it registered the bullish macro signal correctly. What it couldn't do was translate that macro read into a precise intra-window position.

That distinction cost the bot three trades it should have owned.

InDecision's six factors — Daily Pattern, Volume, Timeframe Alignment, Technical Confluence, Market Timing, and Risk Context — are built for directional bias over hours and days. They answer the question: does this asset have a confirmed direction? What they don't answer is: within this specific 15-minute window, how far has price already moved from the open, and which direction is it still running? Those are different questions operating on a different timescale. The model was answering the first one correctly while the second one went unmeasured entirely.

Three Trades, Three Mechanisms

Trade 1: BTC UP, 11:45AM–12:00PM ET. BTC was running on the +6.67% daily candle. The 1-hour context was bullish across the entire market — BTC, ETH, SOL, and XRP were all showing 100% UP simultaneously on the Polymarket 1h contracts. This is not a subtle signal. Four correlated assets with full directional alignment on the 1-hour timeframe is a structural read, not noise.

I bought UP at a 74¢ average across 14 shares. Total cost: $9.90. The window closed UP. Net gain: +$3.75, a 37.87% return on the position.

Trade 2: BTC DOWN, 12:00–12:15PM ET. The daily trend was still bullish — BTC was +6.67% on the day. The obvious read from daily structure was "stay long." That's the read the bot had. The intra-window read was different.

BTC opened the 12:00 window at $68,301. By the time I was evaluating the trade, it had already moved to $68,095 — a −$206 move, −0.30% from window open. That delta was visible in the price feed. The market was pricing DOWN at 53¢, implying roughly 47% probability of a down close. My read, based on the intra-window trajectory, was closer to 70%.

Bought DOWN at 53¢, 9 shares, $4.92 cost. Won +$3.48, a 70.79% return. The bot scored the setup as a low-conviction UP (daily trend bias) and skipped. The mechanism it was missing: price relative to window open.

The window-level trajectory was directionally unambiguous. The −$206 move from the window open had already established a clear intra-window trend — BTC wasn't hovering, it was actively running. The delta between what the market implied (47% probability) and what the price trajectory showed (~70%) represented 23 points of mispriced probability. That isn't edge requiring sophisticated analysis. It's edge left on the table by a model looking at the wrong clock.

Trade 3: Predictive UP on a future 15m window. The third trade required a different kind of information entirely — one the bot literally cannot access. While the 12:00 window was still active, Polymarket had the 12:15, 12:30, and 12:45 windows open simultaneously, each at discounted odds. The 12:15 UP token was trading at 43¢. The four-asset 1h alignment was still intact. I bought the future window before it became the current window.

Won. The pre-market discount — 43¢ vs. what UP would have been priced at window open — was the entire edge.

The Two PRs

These weren't post-hoc rationalizations. The gaps were observable in the bot's scoring output — specifically what factors were present vs. absent when it produced a 65/100 and declined.

feat/window-momentum adds a new scoring factor: price delta from window open to current price. If BTC is sitting −0.30% from the open of the current window and the evaluation is scoring a DOWN bet, that's a structural alignment the model should weight. Score range: −15 points (window moving against the thesis) to +20 points (window open moving in the thesis direction). This is the mechanism that would have caught Trade 2.

feat/predictive-stacking addresses the discovery problem. The bot currently only resolves the active window. Polymarket exposes up to three future 15-minute windows simultaneously, each at discounted pre-market odds. The PR adds a secondary resolution pass: when 1-hour trend context scores BULLISH above a confidence threshold, the bot evaluates future windows with an odds gate (skip if already priced >55¢). Trade 3 was 43¢. That passes the gate.

What a Human Knows That the Model Doesn't (Yet)

The 65/100 score wasn't wrong given what the model could see. The macro signal was real. The daily trend was real. On daily-timeframe data alone, a 65 is defensible — not enough conviction, skip.

What broke the model was context it wasn't collecting: the window's own price trajectory from its open, and the existence of future windows sitting at pre-market discounts. Neither of those inputs existed in the feature set. The bot scored off information it had. I scored off information it didn't have.

That's the actual lesson from today. Not that the bot is insufficient, but that every human override that produces a win is a labeled training example for which features were missing. Three wins today means three mechanisms to encode. The model gap was specific, measurable, and fixable — and it took watching the bot sit on three live setups to surface the exact shape of the blind spot.

PolyEdge v4.1 runs with 1063 tests and 93% coverage. Both PRs ship with full regression test suites — window-momentum with 29 new tests covering aligned/opposed moves across magnitude bands and odds discount gates, predictive-stacking adding coverage for the upcoming-window discovery path and the 55¢ odds gate. The coverage floor isn't bureaucratic overhead. It's the mechanism that makes iterating on a live trading system safe — every change is bounded, every edge case is named, and every regression is caught before it costs real money.

The model learns the same way every other model does: by being wrong in ways that are precise enough to correct.

The model learns the same way every other model does: by being wrong in ways that are precise enough to correct.