Cross-venue calibration: Manifold vs Kalshi on the same Fed events

Why this comparison matters

Calibration is the question every prediction market eventually faces: "when the market says 70%, does it happen 70% of the time?"

For most categories the empirical answer hovers near "yes, but with bias". The classic finding (well-replicated since the 1970s on parimutuel betting) is the favorite-longshot bias: longshots (markets pricing <10%) systematically over-resolve YES; favorites (markets pricing >90%) under-resolve YES. The gap is real money on real-money venues; on play-money venues it's smaller, sometimes absent.

A 2026 Fed Reserve paper (covered in Axios, Feb 2026) gave Kalshi a "perfect forecast record" on Fed rate decisions the day before each FOMC meeting — beating fed funds futures. That's a strong claim. The natural follow-up: does Manifold beat Kalshi on the same events?

The dataset surface

Both venues run identifiable Fed event markets — typically titled "Will the FOMC raise/hold/cut rates by X bps at the Y meeting?". Manifold has the wider category coverage; Kalshi has tighter contract specifications and more volume. The cross-venue match is at the level of same underlying event, slightly different question wording.

Our canonical schema makes this tractable: both venues map into the same markets + outcomes tables with a category, resolution_value, and a resolved_at timestamp. The matching problem becomes a fuzzy text-similarity join on the question text plus a closes_at alignment within ±48 hours.

WITH kalshi_fed AS (
  SELECT market_id, title, closes_at, resolved_at,
         resolution_value ->> 'YES' AS yes_payout
  FROM markets
  WHERE venue_id = 'kalshi'
    AND title ILIKE '%FOMC%'
    AND resolved_at IS NOT NULL
),
manifold_fed AS (
  SELECT market_id, title, closes_at, resolved_at,
         resolution_value ->> 'YES' AS yes_payout
  FROM markets
  WHERE venue_id = 'manifold'
    AND (title ILIKE '%FOMC%' OR title ILIKE '%fed%rate%')
    AND resolved_at IS NOT NULL
)
SELECT k.title AS kalshi_title, m.title AS manifold_title,
       k.closes_at, k.yes_payout, m.yes_payout AS manifold_yes_payout
FROM kalshi_fed k JOIN manifold_fed m
  ON ABS(EXTRACT(epoch FROM k.closes_at - m.closes_at)) < 172800;

In Phase-0 we have a 50-market sample per venue, which is too small for proper Fed coverage. The example query is the shape, not a quantitative claim. Phase 1 expands the Manifold + Kalshi backfills to cover 2023–2025 macro events fully, which is roughly ~80 paired Fed-decision contracts to cross-check.

The expected pattern

Three calibration plots become available once the matched pairs exist:

1. Implied probability vs realized. Bin Kalshi's 24-hour-pre-close implied probability into deciles. For each decile, plot the empirical YES rate. A perfectly calibrated venue produces the diagonal. Favorite-longshot bias bows away from the diagonal: high-probability deciles fall below, low-probability deciles rise above.

2. Cross-venue spread. For each matched pair, compute p_kalshi − p_manifold at T-24h. Distributional summary by category (Fed, election, sports). If Manifold is consistently below Kalshi on FOMC contracts, that's a structural belief difference.

3. Basis decay. Track the spread's evolution from T-7d to T-1h to T+0. Does it converge as resolution nears, or diverge? A converging spread with Manifold consistently below suggests Manifold is the "smart" venue (real-money traders should arbitrage in the Manifold direction). Diverging suggests one venue's pricing is capturing information the other isn't.

The arb (back-of-envelope)

Real-money arb across Kalshi and Manifold is structurally hard:

• Manifold is play-money. There's no real-money payout, so positions don't hedge Kalshi exposure in dollar terms.
• Kalshi requires KYC + a US bank account. Cross-venue automation requires both accounts open and funded.
• Position size on Manifold is small (~$1,000 mana = ~$0 USD).

But the signal arb is real and works for institutional users:

1. Detect when Manifold significantly disagrees with Kalshi on a Fed event (e.g., Manifold 35% YES, Kalshi 60% YES).
2. If Manifold's track record is calibrated and Kalshi shows favorite-longshot drift, Kalshi's 60% is an over-confident favorite. Take NO on Kalshi at $0.40, expected edge if Manifold is right ~25%.
3. Position-size by the calibration confidence interval, not the raw spread.

For a pair-trading desk, this is a residual risk strategy on top of fed-funds futures. For a calibration researcher, it's the empirical question of which venue is "right" on what.

What our dataset enables

Two entries into this analysis:

Backtest mode. Pull all matched (Kalshi, Manifold) pairs with resolved_at IS NOT NULL. Compute calibration plots per venue and the cross-venue residual. Reproducible because both venues' resolutions are stored in outcomes.final_payout.

Live signal mode. Pull active matched pairs (both resolved_at IS NULL). Subscribe to price updates on both venues. Alert when |p_kalshi − p_manifold| > threshold with the threshold calibrated against the historical spread distribution.

The Phase-1 expansion adds 2023–2025 macro coverage and a published cross-venue reconciliation matrix — i.e., for every Kalshi market, the closest Manifold equivalent and the historical spread distribution. That artifact is the closest thing to a "calibrated reference" we know how to build.

Caveats

• Title matching is hard. Manifold users phrase questions creatively; Kalshi titles are standardized. Phase 1 includes an embedding-based matcher (we use voyage-finance-2 in a parallel project; same approach here) with manual verification on the top 100 paired markets.
• Resolution definitions diverge. Manifold "Will Powell announce a 25 bps cut at the May meeting?" might resolve YES on a 25-bp cut OR a 50-bp cut depending on creator intent. Kalshi's contract spec is precise. We use Kalshi's spec as the canonical interpretation when reconciling.
• Liquidity asymmetry. Kalshi has 100–1000× the dollar volume on macro events; Manifold has more user diversity but tighter spreads on Manifold- specific markets. Calibration robustness suffers when either side is thinly traded.

Takeaways

1. Cross-venue calibration is one of the most useful research outputs the dataset can produce. It tells you which venue to trust on what.
2. The matching is fuzzy but tractable with a normalized schema + embedding lookup.
3. For institutional traders, the spread is a signal, not an arb — Manifold's play-money limits direct hedging, but disagreement encodes information.
4. Phase 1 ships the matched-pair artifact. Email us if you're working on a calibration paper and want a preview of the matching code.

Email for a preview of the matched-pair dataset, or read the schema to see how the venue fields line up.