Prediction-market fees ate a profitable strategy

Jupiter Pred MKT is a paper-trading engine that runs eight strategy sleeves across Kalshi and Polymarket — divergence between venues, arbitrage locks, whale following, news drift, fair-value edge, favorite–longshot bias, attention scalping and calendar-ladder structure. On 22 September 2026 we audited every closed position since launch. The headline is not that the strategies were wrong. It is that fees turned a winning book into a losing one.

Fifty-nine trades, one problem

Before fees the engine made money. After fees it lost about 1% of the book. Fees were one and a half times the gross profit. Three mechanisms did the damage, and all three are general to prediction markets, not specific to this engine.

Penny contracts are the most expensive contracts

Kalshi's fee is a per-contract charge of roughly 7% × price × (1 − price). That curve is nearly flat as a share of the money at risk when the price is low: at 1.5¢ the fee is about 7% of notional per side, at 16.5¢ about 5.8%, at 50¢ 3.5%, at 90¢ under 1%. A strategy that likes cheap contracts is, without knowing it, a strategy that pays the highest fee rate on the exchange.

Four positions in the ledger were entered below 2¢. Together they paid $8,377 in modeled fees. The single worst was a $40,116 position on a 1.95¢ Polymarket contract — 2.06 million contracts — that exited at exactly the entry price. Its entire result, −$4,114, was fee.

Timeout exits pay to close a thesis that would have resolved for free

Fifty of the fifty-nine exits were timeouts: a hold window expired and the engine sold at the mid regardless of thesis. Sixteen of those closed at the identical price they were bought at, so the trade's only outcome was paying fees twice. Settlement, by contrast, is free — a contract that resolves pays $1 or $0 with no exit fee. An engine that dumps at the mid two days before resolution is paying to avoid the one exit that costs nothing.

Sizing that ignored the fee it was about to pay

No sleeve compared its expected edge to the round-trip fee before entering. The divergence sleeve entered on a 3¢ gap between venues and exited when the gap closed to 1¢, targeting 2¢ of capture; on a 16.5¢ Kalshi contract the modeled round trip is 1.9¢. The arbitrage sleeve locked a 3.5¢ gross spread with legs whose combined fees exceeded 4¢. Both were structurally unprofitable at entry and no rule caught it.

Three rules, all shipped the same day

1. Price floor. No entry below 8¢ on either venue.
2. Fee gate. Every strategy models the round-trip fee before entering. Pair strategies must show an edge larger than the fee plus a margin; sensor strategies may not enter where the round trip exceeds 4% of notional.
3. Hold to settlement. Markets past their close are asked for their result and positions settle at $0 or $1 with no exit fee. Arbitrage locks are never dumped at the mid, and are only taken on markets that settle within 120 days.

The first run after the change settled four open positions at the venues' actual results; all four paid out, +$7,066 combined. That is one run and proves nothing about the strategies. It proves only that the exits were the problem.

Three questions for any prediction-market strategy

• At the prices you trade, what share of notional is the round-trip fee? If you do not know the number, you are probably paying the top of the curve.
• What share of your exits are timers rather than theses or settlement? Every timer exit is a fee paid to avoid a free one.
• Does the entry rule see the fee? If edge is compared to zero instead of to cost, the strategy can be right on direction and still lose.