A 97.7% Win Rate That Loses Money: Buying Favorites, Backtested

The most widely repeated Polymarket Up/Down strategy is also the simplest: with a minute left, buy whichever side is clearly winning. It is intuitive, it is easy to automate, and it produces the kind of win rate that gets screenshotted.

We tested it against 17,002 simulated trades on real order book snapshots. The win rate is genuinely extraordinary. The returns are negative.

How the test works

• Universe: every resolved BTC, ETH and SOL 5-minute Up/Down market in the PolyTest archive — a rolling 31-day window covering 5 August to 5 September 2026.
• Entry: one trade per market, at the snapshot nearest 60 seconds before expiry. One observation per market, so no autocorrelation from sampling the same market repeatedly.
• Side: whichever outcome is quoted higher — the favourite.
• Fill price: the ask, not the mid. price_up + price_down averages 1.0117 in this data, so the roughly 1.2¢ round-trip spread is already paid at entry.
• Exclusions: markets where either side has an empty ask book. If nobody is offering, the trade is impossible — including it would fabricate fills that could not have happened.
• Payout: winners settle at $1.00, losers at $0.

Result: buying the favourite, by price band

Read the bottom two rows first, because that is where the volume is. Over half the trades fall at 90¢ or above, and both of those bands lose money at high statistical confidence. The 90–97¢ band is the worst: you pay 93.56¢ for something that happens 91.16% of the time, bleeding 2.39 percentage points on every trade with a z-score of −5.09.

The middle bands look better — 60–70¢ shows +3.16% ROI — but neither clears significance. At z = 1.66 and z = 1.49 those are the returns you would expect from a coin flip dressed up in 1,500 trades. No band shows a statistically significant positive edge.

The calibration curve

Stepping back from the strategy to the market itself: how accurate are these prices? For every BTC 5-minute market we took the quoted Up price 60 seconds before expiry and compared it to how often Up actually won.

This is a textbook favourite–longshot bias, the effect documented in racetrack betting and prediction markets for decades. Longshots are systematically overpriced: contracts quoted around 17¢ win 13.2% of the time, and ones quoted around 22¢ win 18.0%. Mid-range favourites are underpriced by a few points.

It is real, it is in the right direction, and it is very hard to monetise — because the bands where the bias is largest are the bands with the thinnest books and the widest relative spreads. The bias shows up in the calibration curve and disappears in the P&L, which is the usual fate of a well-known anomaly.

So does the bias exist or not?

Both, and the distinction is the whole point:

• As a description of prices — yes. Longshots below ~25¢ are consistently overpriced by 3–4 percentage points. That is a genuine, measurable property of these markets.
• As a strategy — no. Buying favourites returns nothing after the spread, and at the prices where most volume sits it returns less than nothing with high confidence.
• The asymmetry is the tell. If longshots are overpriced, the trade is not to buy favourites — it is to sell longshots. That requires posting resting offers and being filled by someone else's market order, which is market making, with all the inventory risk that implies. It is not a strategy you execute by hitting asks.

Reproduce it

Every figure above comes from data you can pull. The shape of the test matters more than our numbers — run it on your own window and see whether it still holds.

import os, requests
from statistics import mean

BASE, H = "https://api.polytest.io/api/v1", {"X-API-Key": os.environ["POLYTEST_KEY"]}

trades = []
markets = requests.get(f"{BASE}/markets", headers=H, params={
    "coin": "btc", "market_type": "5m", "limit": 100,
}).json()["markets"]

for m in markets:
    if not m.get("winner"):
        continue
    # The snapshot nearest 60s before expiry — one observation per market.
    snap = requests.get(f"{BASE}/markets/{m['id']}/snapshots/at", headers=H, params={
        "timestamp": m["end_time"], "offset_seconds": -60,
        "include_orderbook": "true",
    }).json().get("snapshot")
    if not snap:
        continue

    up, down = snap.get("price_up"), snap.get("price_down")
    if up is None or down is None:
        continue                       # no offer on one side: untradeable

    side = "Up" if up >= down else "Down"
    cost = max(up, down)               # pay the ask, not the mid
    trades.append((cost, m["winner"] == side))

n = len(trades)
cost = mean(c for c, _ in trades)
wr = sum(w for _, w in trades) / n
print(f"n={n}  cost={cost:.4f}  win_rate={wr:.4f}  ROI={(wr-cost)/cost*100:+.2f}%")

What this does and does not prove

• One month. A rolling 31-day window, 5 Aug – 5 Sep 2026. Long enough for 17,002 trades, not long enough to cover multiple volatility regimes.
• Top-of-book fills. We charge the best ask and assume the full trade fills there. Real size walks the book, so live results would be worse than these, not better.
• No fees or gas. Adding them moves every row down.
• 5-minute markets only. Longer timeframes have different liquidity and may behave differently.
• Entry at exactly T−60s. Timing turns out to matter enormously — see the entry timing study.

The conclusion we would defend: buying favourites at the ask is not a strategy, it is a way to pay the spread 17,000 times. If you want a version of this that does work, the spot-lag study tests a related idea that does show a real edge — along with the constraint that stops it scaling.

Every number here came from PolyTest snapshots with full order book depth. Polymarket's own API cannot support this test at all: request a resolved market's book and it returns 404, request its price history and it returns 200 with an empty array. Start free if you want to run your own version.