KPI Edge · mechanism

From a probability ladder to a stock-moving surprise

One name, real Kalshi data, end to end: Tesla's Q2 production market — how its probability ladder becomes an implied distribution, and the first honest test of whether the change in that implied number tracks the stock.

01The raw market — a ladder of "above X" contracts

KXTSLA-26JULPROD · "Tesla Q2 2026 production" · closes 2026-08-21 (at the production report).

Kalshi lists one binary contract per threshold. Each YES price is the market's probability that the reported figure lands above that level — i.e. a point on the survival curve of the metric. Tesla's ladder is deep enough to be clean and monotone: 8 brackets from 320k to 460k vehicles.

Contract (threshold)	Implied P(report > threshold)	Reads as

02Differencing the ladder → an implied distribution

Subtract adjacent survival probabilities and you recover the implied probability mass in each band.

Implied probability mass (band) Survival curve P(> X)

Derived from live KXTSLA-26JULPROD YES prices, 2026-06-15.

≈ 430k

Implied consensus (median — where P(>X) crosses 50%)

~25%

Implied mass in the 420k–440k modal band

left-skew

Long lower support (94%+ above 380k), thinner upside

This is the payload a point estimate can't give you: a shape. You can read the median as the "whisper number," the modal band as conviction, and the skew as the asymmetry into the print.

03The real test — does the change in implied consensus track the stock?

An event date by itself tells you nothing. The signal — if it exists — is in the daily drift of the implied number vs. the equity.

Tesla's Q2 production market has ~60 days of daily traded history across its 8 brackets — enough to actually test this (most KPI markets are still too new). Reconstructing the implied consensus production (the ladder median) each day and overlaying TSLA's close gives the first honest look at the relationship.

Implied consensus production (ladder median, left axis) TSLA close (right axis)

Live KXTSLA-26JULPROD candlesticks + TSLA via yfinance, 40 aligned trading days to 2026-06-12.

What the correlation actually is

Relationship	Pearson r	Read
Level — implied production vs. price	−0.19	weak / noise
Change — Δimplied vs. same-day return	+0.08	≈ zero
Lead — Δimplied (t) vs. return (t+1)	+0.09	≈ zero

Honest verdict (n = 40): for this one name and one quarter, the drift in implied production shows no meaningful correlation with TSLA — every r is well inside the noise band (you'd need |r| > ~0.31 to call it real at this sample size). That is a genuine result, not a rendering of a thesis. Three readings, and the product hinges on which is true:

Day-to-day production drift just isn't the stock's marginal driver — the KPI may only matter at the print, as a surprise, not as a slow drift. (That sends you to the surprise study below, not the drift.)
A single bracket-derived median is too noisy on thin volume to carry signal day-to-day.
n = 1 name × 1 quarter proves nothing — which is the whole point: the only way to know is the pooled study across ~70 names and many quarters. One chart can't settle it; the dataset can.

The surprise study this motivates

The drift view above is the leading-indicator question. The complementary, likely-stronger test is the surprise at resolution — which the repo already computes (surprise = resolved − pre_event_prob, joined to returns via horizon_return()). Swap the macro ETF basket for the single ticker and align on the report date:

row = {
  "name": "TSLA",
  "event_date": "2026-08-21",
  "implied_consensus": 430_000,      # median of the implied ladder
  "actual": q2_production,           # reported figure (settlement source)
  "surprise": actual - 430_000,      # the signal
  "tsla_ret_1d": horizon_return(rets, "TSLA", event_date, 1),
  "tsla_ret_5d": horizon_return(rets, "TSLA", event_date, 5),
}

Pool these rows across ~70 names and many quarters and you have the core test: does implied-vs-realized surprise predict the event-window equity return? Plus the pre-print drift signal — the daily change in implied consensus — as a leading indicator before resolution.

04The pipeline that produces it

Most of this exists today — it was built for the Fed/macro lane and ports directly to single names.

Kalshi API→ archive ladder (daily)→ implied-prob panel→ consensus + drift + surprise→ join equity returns→ signal / backtest

kalshi_core

REST client · models · candlestick history · archival

pm_research

prob panels · daily snapshots · calibration (Brier / log-loss)

finmap

ontology · equity joins · event study · surprise

What's new here: finmap/drift_study.py reconstructs each name's implied consensus from candlesticks and runs the pooled Δconsensus-vs-return test (kalshi-build-drift-study) — plus a liquidity / coverage screen so we research the names that actually trade.