Polymarket BTC Up/Down 5-Min Markets: Walk-the-Book Reality Check

If you've built a backtest for Polymarket's BTC Up/Down 5-minute markets and the equity curve looks too good, here's the most likely reason: you priced fills at the displayed (best_bid + best_ask) / 2 mid. For the first 4 minutes of a 5-minute market's life, that number is a fiction. The two-sided book is typically 11¢ bid / 89¢ ask, the mid is ~$0.50, and nothing fills at $0.50. Any non-trivial order has to walk the book.

We've been writing every level-2 BTC orderbook update to ClickHouse since early 2026. This post is what that data says about the gap between simulated and realisable PnL.

The book shape, with numbers

Here's a typical first-4-minutes snapshot, taken from the actual market id our tracker logged earlier today:

The book stays this thin until roughly τ < 60s. At that point, taker flow picks up and the spread typically compresses to 50¢ / 51¢ as market-makers pull risk and informed traders take it. The compression itself is the alpha window — but it's a 60-second window, not a 5-minute one.

What this does to a textbook strategy

Take a vanilla "buy YES when YES probability > 0.6" backtest. Run it at the displayed mid, you get cumulative PnL that looks fantastic. Run it at the walk-the-book fill, you pay $0.89 for shares that pay $1.00 if you're right and $0.00 if you're wrong — an 11¢ edge before fees, ~7¢ after Polymarket's 2% taker. Your win rate has to clear 88% just to break even.

That's the entire reason walk-the-book fills + fee modeling are the default in PolyQuantLab's backtest engine, not an option. Showing pretty PnL from mid-price simulations is the marketing trick we explicitly designed against.

What works on BTC 5m markets

Three patterns survive walk-book + fee accounting:

1. Logical arbs only. When YES ask + NO ask drops below $1 in walk-book terms (rare, but it happens 5–15 times per day across all BTC/ETH/SOL markets), buy both. See the arb audit.
2. Endgame model edges in the last τ < 3 min. When realised vol on Binance is low and our log-normal probability disagrees with the market by > 15 points, the edge is large enough to clear the spread. Limited frequency: ~7 actionable signals/hour across all coins/timeframes.
3. Maker-only quoting at the inside. If you have queue priority and the spread compression actually fires, your average cost is closer to the mid than to the ask. Phase Z of the backtest engine models queue position; without it, assume your maker orders never fill ahead of the MM.

What doesn't work

• Mid-price entry with mid-price exit. If you paid the mid you would have run away with it. You didn't.
• Momentum on the Polymarket order flow alone.The order flow lags the Binance flow by 30–90s. Any signal you derive from it has to compete with the model that already incorporates the leading Binance data.
• Threshold entry at > 0.5 YES with no model.Buying a coin-flip at 0.6 with no edge over the market is paying a 10¢ premium for noise.

How to reproduce this

1. Sign up (free tier) at polyquantlab.com/signin.
2. Open /dashboard/markets and filter to ticker=BTC, event_type=5m. Click any market.
3. Watch the orderbook timeline. You'll see the 11¢/89¢ pattern hold for ~80% of the market's life.
4. Click "Use as backtest universe" and toggle fill_mode between "mid" and "walk-book" on the same strategy. The PnL delta is the gap we've been describing.