Trading 15-minute Polymarket markets until the data said stop

After pausing long-horizon prediction in part 1, I moved to Polymarket's 15-minute Up/Down markets. Each one asks whether an asset such as SOL finishes a 15-minute window up or down. Up and Down shares trade between $0 and $1, and the side that turns out right pays $1 at expiry.

The attraction was the loop: faster iteration, cleaner labels and a short path from research to execution. The hypothesis was that such a short expiry might still leave room for timing mistakes or order-book inefficiencies.

The plan: record snapshots continuously, enrich every market with Binance volatility context, sweep large parameter grids across four entry modes, then run a live system with real orders and real fills.

The research layer compared four ways to enter a market. FADE and MOMENTUM were the two core reactions to early price behavior; REVERSE and DUAL_ENTRY were alternative entries for noisier setups.

Each market was split into blocks, and each block carried its own absolute prices for entry, take-profit and stop (entry_range, tp_range, stop_range) instead of relative deviations. Vectorized backtests over these block configurations made large 3D parameter sweeps practical.

Volatility changed what a good configuration looked like. For every market, volatility_by_slug.csv stored Binance volatility over 15 minutes, 1 hour and 4 hours, plus a low, medium or high regime label. Rules could then differ by regime instead of relying on one global setting.

Selection was walk-forward: parameters chosen on one period were judged on the next, with diversity across volatility regimes built into the choice. An optimizer wrote the winner to best_config.json, along with an optimized rules file.

A flip analyzer asked a narrower question: when is a very cheap price late in a market effectively dead, rather than a bargain waiting to flip?

The biggest improvement in this phase was not a new strategy. It was making the simulator stop flattering me.

Three rules changed everything. Inside a block, orders are processed in a fixed sequence (stop, then take-profit, then entry), so a position cannot be opened and closed on the same future information. Every decision waits out a reaction latency before it can trade. And every fill pays for execution through fees, slippage and a liquidity filter.

The base simulator carried realistic frictions. Stress runs pushed much harder, to see whether an edge survived pain.

The live bot placed limit orders on Polymarket's central limit order book (CLOB) from a server, so it could run around the clock. It made one thing clear very quickly: execution logic is mostly synchronization logic.

WebSockets had to be the source of truth for prices and fills, with roughly 50 ms responsiveness against about 500 ms for REST-style polling. REST stayed useful for reconciliation, not for pretending to be real time. Open orders and real inventory had to be tracked separately.

Partial fills, ghost orders and the rollover from one 15-minute market to the next were normal conditions, not exceptions. The general lessons are in The reality of bots on Polymarket; the loop below is the one this bot ran.

That conclusion was disappointing emotionally and excellent technically. The system had become honest enough to reject the thesis instead of flattering it.

The live bot's hardest problems were synchronization, not entry and exit logic. But it did not stop because the code was broken. It stopped because the premise of an edge was wrong for this market.

Stopping when the data says there is no edge is a result. It only counts as one because the measurement had been made hard to fool.

This phase did what a research phase should do: it reduced uncertainty with a system strong enough to say no. It did not leave me a strategy worth running. It left a repeatable proof that the easy edge was not there.

The next question was narrower: could second-scale signals still hold something the 15-minute window did not? Answering it meant recording my own data, moving the hot path to Rust and treating timestamps as seriously as fills. That is part 3, HFT on Polymarket.