After the 60s TWAP Switch: Live Lessons From Running a BTC 5m Polymarket Bot

On August 14, 2026, Polymarket altered the 5-minute crypto Up/Down markets from a 30-second Chainlink Time Weighted Average Price (TWAP) to a 60-second TWAP. A specialized BTC 5-minute directional engine was directly impacted by this modification. This report details the actual changes in live operation following the cut-over.

The first 48 hours post-cut-over felt quieter than anticipated. The most noticeable difference was the reduction of late-slot spikes that used to be significant, as they were averaged out in the new 60-second TWAP. The settlement path appeared less dramatic on the chart. Some "obvious" last-minute flips ceased to resolve in the same manner, which aligns with the expected behavior of a longer lookback period. This change also meant that logic relying on quick fluctuations had less to work with.

The directional strategy, which involved forecasting the official TWAP path, entering mid-band, and discarding weak tickets, remained coherent. What changed was the magnitude of features, the timing of useful edge shifted to earlier or mid-slots for certain scenarios, and scratch decisions became the primary performance lever. In essence, the model still had insights to offer but the old calibration was no longer effective.

A dangerous failure mode that became silent after the settlement change was mismatch. The worst bug post-settlement change isn't a crash but a bot that runs while addressing the wrong issue. Signs to watch for include paper or internal scores drifting from gamma resolution, open references pinned from a leftover 30-second path, 'win rates' appearing fine in local logs while live redemption disagreed, and conviction signals firing on noise that no longer impacts payout.

If monitoring systems only track fills and not resolution source plus TWAP window, this type of error can go unnoticed for days.

Switching the real-time decision service (RTDS) topic to 60-second TWAP was relatively easy, but managing position culling proved challenging. Old decay thresholds were overly sensitive to the smoother series, turning some previously valid scratches into unnecessary churn. Some holds that seemed fine under 30-second rules turned out to be clear losers under the new 60-second settlement.

The scratch layer is where short-horizon expectancy is created or destroyed. After a window change, the scratch layer needs to be retuned before trusting size.

Practical adjustments that aided the transition included explicit window tagging in every decision log, a shadow comparison scoring both legacy 30-second and live 60-second views, reduced size until stability was achieved, hard skipping of stale TWAP series, and separating research from production. Research can still analyze 30-second data for historical comparisons, while the production 5-minute path is limited to the 60-second TWAP only.

Several improvements were observed post-upgrade: fewer outcome flips driven by brief price movements, a cleaner relationship between the "path so far" and final settlement, better integrity for those forecasting the official average, and a heightened need for honest labeling in paper trading. Architecturally, the emphasis on settlement alignment over clever features proved essential.

If the target variable changes, features are secondary until the target is correctly identified. Specialization also helps; a bot focused on BTC 5-minute markets can adapt faster than a multi-market system with hidden assumptions. Observability is crucial for preserving alpha, as you cannot improve what you cannot reconstruct after the fact.

Scratch strategies require re-estimation and should be treated as first-class components, not just execution details. Configurability is vital; parameters like TWAP window, staleness limits, and resolution source should not be hard-coded.

Currently, the 5-minute crypto path is fully running on the 60-second Chainlink TWAP. The open reference and final score are sourced from the same series. Gamma resolution is used as the authoritative resolution when present. Scratch thresholds are being actively re-estimated, and size is gated until post-cutover metrics stabilize. The system retains its high-level design: TWAP-anchored directional read, mid-band entries, and active culling, but now tailored to the actual settlement process Polymarket employs.

In summary, the 60-second TWAP update did not dismantle short-horizon directional trading on Polymarket. Instead, it highlighted the importance of updating assumptions. Bots that merely changed a subscription string without retuning signals, labels, and scratch behavior would typically experience quiet degradation. Conversely, bots that treated the cut-over as a comprehensive target-variable migration could maintain alignment.

If you're operating anything on BTC 5-minute markets (or other similar markets), verify three things today: Are you consuming the 60-second series? Are open and settlement pinned to that same series? Are your live results reconciling cleanly against gamma resolution? Everything else is optimization.