Why 68% Accuracy Lost Money: Payoff Asymmetry in Prediction Markets

We built a weather prediction model that was right 68% of the time. It lost money.

Not a little money. Every price bucket we tested showed negative P&L despite win rates that would make most forecasters proud. The model had genuine skill — it outperformed both random chance and the market's implied probabilities at predicting temperature outcomes. And it still couldn't generate a profit.

The reason is payoff asymmetry, and if you're building any prediction market strategy, understanding it will save you from the most expensive mistake in the space.

The Setup

We've been running a weather prediction system against Kalshi's temperature markets. Binary contracts: "Will Denver's high exceed 58°F tomorrow?" pays $1 if yes, $0 if no. We use NOAA forecasts to estimate probabilities and compare against market prices.

After filtering for real liquidity, we ran a bucket analysis across all settled contracts. The model showed a clear strength: it's much better at predicting what won't happen than what will.

68% on the NO side. The model genuinely knows when something isn't going to happen. If NOAA says 45°F and the market asks "will it exceed 60°F?", our model correctly says "no" about two-thirds of the time.

So we should trade the NO side, right? Buy NO contracts where the model is confident. Simple.

Here's where it breaks.

The Payoff Trap

When the model says "this won't happen" on a weather contract, it's usually because the forecast is far from the strike price. NOAA says 45°F, strike is 60°F — obviously not happening.

But the market knows this too. These contracts are priced accordingly:

See the problem? The contracts where our model is most confident (far from the strike) are also the ones with the worst payoff ratio. Paying 90¢ to win 10¢ means you need to be right 90% of the time just to break even.

Our model is right 68% of the time. That's good. It's not 90% good.

The Breakeven Math

For any binary contract, the breakeven win rate depends entirely on the entry cost:

The formula is simple: Breakeven = Cost / (Cost + Payout), which for binary contracts simplifies to Breakeven = Cost / $1.

A 68% accurate model is profitable at entry costs below 68¢. Above that, the math works against you no matter how good your predictions are.

Most of our NO-side trades were in the 80-95¢ range. We were paying 85¢ to win 15¢, needing 85%+ accuracy where we had 68%. Every trade had negative expected value despite the model being "right."

What the Buckets Showed

We sliced the data by entry cost to find where, if anywhere, the economics worked:

One band — 40-60¢ entry cost — showed 80% accuracy and positive P&L. Everything else lost money.

Why the 40-60¢ Band Works

The 40-60¢ entry cost band is special for two reasons:

1. Balanced risk/reward. At 50¢, a win pays 50¢ and a loss costs 50¢. You only need 50%+ accuracy to be profitable. At 55¢, you need 55%. Our model delivers 80% in this band — well above breakeven.

2. Genuine disagreement zone. Contracts priced at 40-60¢ are ones where the market is uncertain. The market is saying "this could go either way." But if our model confidently says NO with 80% accuracy, that's a real information advantage — we know something the market is unsure about.

Compare this to the 90¢ band: the market already agrees with us (pricing YES at only 10¢). There's no informational edge — we're just paying a premium to bet on something everyone already expects.

Why the Model Breaks Below 40¢

The 20-40¢ band went 0 for 10. Why?

Contracts in this price range for NO-side trades mean the market prices YES at 60-80¢. The market thinks the event is likely. Our model disagrees. But in this range, we're fighting against genuine probability — these events actually DO happen more often than our model thinks.

This is the thin-tail problem we've documented before. Our model uses a Gaussian (normal) distribution to estimate temperature probabilities. The normal distribution underestimates the frequency of large forecast errors — "fat tail" events. When the model says "only 20% chance of exceeding 55°F" but reality has fatter tails, the market's 65% price is closer to truth than our 20%.

The model is well-calibrated in the middle (40-60¢ range) but poorly calibrated at the extremes (both tails). This is a known limitation of Gaussian models and fixable with better distributional assumptions — but it defines where we can and can't trade today.

Implications for Any Prediction Market Strategy

This isn't unique to weather markets. The payoff asymmetry trap applies to any binary prediction market:

• Political markets: Buying "Biden wins California" at 95¢ requires near-certainty to profit. You're paying for something everyone already knows.
• Sports markets: Betting "Lakers beat the G-League team" at 92¢ has terrible expected value even if they win 95% of the time.
• Crypto markets: "BTC above $10K" at 98¢ — you're paying 98¢ to win 2¢. One black swan wipes out 50 wins.

The profitable zone is always in the middle — where the market is genuinely uncertain (40-60¢ range) and your model has a real informational edge. At the extremes, even accurate models can't overcome the payoff math.

Our Updated Trading Rules

Coming out of this analysis, we added three hard constraints:

1. Never enter above 65¢. Our model's 68% accuracy means anything above 68¢ has negative expected value. We add a 3¢ buffer for fees and model uncertainty.
2. Sweet spot is 40-60¢. This is where our accuracy (80%) far exceeds breakeven (40-60%). Prioritize contracts in this band.
3. Skip the extremes entirely. Below 30¢ our model breaks down (fat tail problem). Above 70¢ the payoff is too asymmetric. Both tails are unprofitable.

The Caveat: 20 Trades

We'd be irresponsible not to flag this: 20 trades is a very small sample. An 80% win rate on 20 trades has wide confidence intervals. We could be seeing signal or noise.

To reach 95% confidence that our true win rate exceeds 60% (our minimum for profitability in this band), we need roughly 35-50 trades in the pocket. We're collecting data through the weekend and will re-evaluate Tuesday.

This is why we haven't risked any capital. Phase 0 is about proving edge exists with statistical significance before money goes in. The pocket is promising, not proven.

The Takeaway

Accuracy is necessary but not sufficient. A prediction market strategy needs three things working simultaneously:

1. Directional accuracy — your model needs to be right more often than not
2. Favorable payoff ratio — the cost of being right must be low enough relative to the reward
3. Sufficient volume — enough contracts in the profitable zone to make it worth running

Most people optimize for #1 and ignore #2. We did exactly that — built a 68% accurate model and watched it lose money because we were trading in the wrong price band.

Don't ask "how often am I right?" Ask "how often am I right, at prices where being right pays?"