Manifold
~7 min read
2026-05-02
The volume field on a Manifold market does not equal
SUM(bet.amount). It also doesn't equal
SUM(abs(bet.amount)), except sometimes. The truth is in the
per-bet fills[] array, and the public /v0/bets
endpoint only surfaces the top level. Here's the reproduction, the
drift quantified, and the workaround.
Polymarket
~9 min read
2026-05-02
The Polymarket data-api /trades endpoint silently
returns 400 above offset 3,500. For markets like Trump 2024 ($1.5B volume,
5M+ trades) you can recover ~0.07% of the real history through REST.
We map the way out: Goldsky's orderbook subgraph for the standard CTFExchange,
plus direct Polygon eth_getLogs for the NegRisk-wrapped markets
the subgraph doesn't index. Code + benchmarks.
Kalshi
~5 min read
2026-05-02
Of the three venues we cover, Kalshi is the cleanest by a wide margin —
50/50 markets reconciled to 0.0% drift on the Phase-0 sample.
What that actually means, how the sample was selected, the v2 API field
rename gotcha (volume_fp vs volume), and why
this matters for academic backtests.
Methodology
~10 min read
2026-05-02
Every prediction-market scraper drops markets when the venue archives them.
The result: backtests run on "current" data are systematically biased
toward strategies that traded markets which never resolved badly.
A 30-line proof, the magnitude of the effect on a 2024 sample, and the
schema-level fix (deletion ledger).
Cross-venue
~8 min read
2026-05-02
Manifold (play-money) is widely cited as well-calibrated; Kalshi (real-money,
CFTC-regulated) shows favorite-longshot bias in some categories. What does
the same FOMC rate decision look like on both venues? Pricing spread,
basis decay, and a back-of-envelope arb. Replication code included.