Intraday Drift Study — Naive NO-GO, Forecast-Conditioned is a Candidate

Research finding from an intraday price-drift backtest on Polymarket weather (max-temp) buckets. The NAIVE drift strategy is dead; a FORECAST-CONDITIONED variant is a live candidate. Buying unlikely buckets blind at open and selling on a +50% pop loses ~27% net because of survivorship bias — you cannot tell, at open, which longshots will moon and which will quietly decay to zero, and the losers dominate. But the Follow-up (v2): forecast-conditioned re-analysis (same 112-market-day cache, no re-fetch) shows the edge is real once you select the bucket via forecast proximity and hold to afternoon convergence: buy-blind ≈ −15% → near-forecast +1% → near-forecast + 18:00 exit +9.6% mean blended return. That is a CANDIDATE, not a confirmed GO (n=125, single month, 80% LCB still −6.9%), and it directly confirms the 2026-06-11 nowcast / information-timing pivot — the edge is picking the right cheap bucket via forecast (and ultimately a live nowcast), not the drift itself.

For Agents

This is a research finding, not code documentation. The NAIVE buy-the-drift strategy is REJECTED (see But the naive strategy is NO-GO, ~−15% to −27% blind). The FORECAST-CONDITIONED variant is a live CANDIDATE (see Follow-up (v2): forecast-conditioned) — +9.6% mean blended return when selecting cheap (10–25¢) buckets at or just above the forecasted high and exiting at 18:00 local convergence, but not yet statistically confirmed (n=125, one month, 80% LCB −6.9%). The two levers — forecast-proximity (winner selection) and exit-timing (hold to convergence) — each help and stack. Liquidity is NOT the blocker (median bucket volume ~$20K,100$1K). The supporting wx-research crate is documented separately in wx-research-crate; the data clients it runs on are wx-market-prices-history and wx-market-trades-client. Market-structure ground truth (11 neg-risk 1°C buckets, ICAO/Wunderground whole-°C resolution) is in polymarket-weather-market-structure.

Question

Does the observed intraday up-drift of unlikely Polymarket weather buckets — the casual observation that “even very unlikely bets drift up during the day” — support a mechanical buy-at-open / sell-before-close-when-up-enough strategy?

• Cities: London, Paris, Seoul, Beijing — daily max-temperature markets.
• Framing: treat each event’s 11 buckets as a discrete distribution (see market structure); ask whether the open-price cohorts that drift up are tradeable blind.

Dataset

• 112 market-days = 28 dates × 4 cities. (May 17–18 absent — no markets existed those days.)
• 1,224 buckets, each with a full intraday price curve.
• Per bucket: lead-0 / lead-1 / lead-2 Open-Meteo previous-runs forecasts + the realized daily high.
• Cached at research/data/; report at research/report.md.
• Code: the new wx-research crate on branch research/intraday-drift (uncommitted at time of writing). Crate skeleton: wx-research-crate. Runs off wx-market-prices-history (CLOB /prices-history) and wx-market-trades-client.

Quick Reference

Re-runs of conditioned variants execute off the existing cache in seconds — no re-fetch needed. Cache: research/data/. Report: research/report.md.

The drift is REAL

By open-price cohort, the fraction of buckets whose sellable peak reached ≥ +50% of entry:

Open-price cohort	Entry band	Fraction reaching ≥ +50% peak
Deep-longshot	(very low)	35.7%
Longshot	(low)	39.4%
Unlikely	10–25¢	49.0%
Live	(mid)	43.1%

• The unlikely (10–25¢) cohort is the sweet spot: ~half reach a ≥ +50% sellable peak at some point intraday.
• Median MFE (max favourable excursion) for the unlikely cohort = 10¢.
• Peaks cluster at 17:00–19:00 local time — late afternoon, when the day’s actual high is being realized and the market converges toward the answer.

So the observation is correct: unlikely buckets genuinely tend to tick up during the day, peaking in the late-afternoon window as the realized high comes in.

But the naive strategy is NO-GO

Pre-registered test (fixed rules, decided before looking at the result):

• Cohort: unlikely buckets only.
• Subset: ultimately-LOST buckets only (the honest population for a blind entry — see mechanism below).
• Exit: fixed +50%, causal (sell the first time price is ≥ entry × 1.5; no look-ahead).
• Costs: applied spread-haircut (Polymarket charges 0 trading fee, so cost = spread only — see Liquidity & volume (observed 2026-06-07)).
• Stats: 80% bootstrap CI.

Result:

Naive buy-open / sell-+50% — REJECTED

Mean net return −27% (80% bootstrap CI −32% to −22%, n = 422). After costs, on the realistic (loser-inclusive) population, the strategy is a clear loser. NO-GO.

Mechanism: survivorship bias

• ~Half the unlikely buckets quietly decay to zero over the day.
• At open you cannot distinguish the future winners from the future losers — they look identical.
• Of the ~11 buckets per market, the longshots that ultimately lose number ~8–9 of 11, so the loser population dominates any blind basket.
• The “drift up” you observe is the survivor path; selecting on it requires hindsight. Trade it blind and the decayers swamp the poppers.

Where the edge likely is

The same +50% causal exit, applied to unlikely buckets that ultimately WIN, returns +54%.

The edge is selection, not drift

Drift is real but un-tradeable blind (−27%). The 100+ pp swing between the loser cohort (−27%) and the winner cohort (+54%) means the entire edge lives in PREDICTING which bucket wins, not in harvesting the intraday drift.

• The obvious predictor is the forecast: buckets near the forecasted high win far more often. A forecast-conditioned filter is the natural way to bias the basket toward the +54% population.
• This is the same thesis as the 2026-06-11 session’s recommended pivot (“intraday nowcast vs stale prices — information-timing edge”): the market does reprice toward the realized high through the afternoon (confirmed here — peaks at 17–19h local), so the opportunity is an information-timing / nowcast edge that picks winners ahead of the crowd, not a blind buy-the-drift trade. This study is the empirical confirmation of that pivot’s premise and the rejection of the naive alternative.

Next step → DONE (see Follow-up v2)

The planned conditioned experiment has been run off the same cache and is captured below. Both levers landed:

1. Forecast-conditioned cohort — restricting to unlikely buckets within ±1°C of the lead-2 forecast flips the blended basket from ~−15% (FAR/blind) to roughly break-even (+1.2%), exactly the predicted shift toward the winner population.
2. Time-based exits — holding to 18:00 local convergence instead of selling the first +50% pop lifts the same NEAR universe to +9.6%.

→ Full numbers, buckets, and caveats in Follow-up (v2): forecast-conditioned.

Follow-up (v2): forecast-conditioned

Re-analysis off the same 112-market-day cache (no re-fetch). The v1 NO-GO killed the naive blind trade; v2 tests the two predicted levers — forecast proximity (winner selection) and exit timing (hold to convergence). Both land, and they stack. The result is a CANDIDATE, not a confirmed GO.

Lever 1 — forecast proximity is a real winner-predictor

Among unlikely (10–25¢) buckets, split by distance between the bucket’s temperature and the lead-2 forecast (|bucket_temp − lead2_forecast|). Returns are after-cost, buy-ALL, +50% causal exit:

Cohort	Filter	n	Win-rate	Blended after-cost return	80% LCB
NEAR	`	dist	≤ 1°C`	125	20.0%
FAR	`	dist	> 1°C`	373	13.7%

• Within the NEAR cohort, won-only = +63.6% (n=25) vs lost-only = −14.2% (n=100) — same survivorship split as v1, but the basket is now tilted enough that blended is roughly break-even instead of −15% to −27%.
• Win-rate vs forecast-distance peaks at dist 0 and +1°C (~21%) and collapses to ~5% at dist ≤ −2°C.

Buy AT or just ABOVE the forecasted high

The buckets that win most are the forecasted high and the one above it (dist 0 and +1). Buckets below the forecast win rarely. Net: forecast proximity is a genuine winner filter — selecting on it moves the blind ~−15% basket to roughly break-even (+1.2%) before any exit-timing improvement.

Lever 2 — exit timing matters more than the threshold

Same NEAR universe (the +1.2% basket above), varying only the exit rule:

Exit rule	Blended after-cost return	80% LCB	80% hi
Sell first +50% pop (causal)	+1.2%	−8.0%	—
Hold to 18:00 local convergence	+9.6%	−6.9%	+26.4%

• Selling the first pop leaves most of the move on the table; the market keeps repricing toward the realized high through the afternoon (consistent with the v1 finding that MFE peaks at 17–19h local).
• Holding to 18:00 nearly 8בs the mean (+1.2% → +9.6%) on the identical entry universe. Exit timing is the bigger lever once selection is fixed.

Liquidity is NOT the blocker

Fills are plausible for this universe

• Unlikely-cohort median bucket volume ~$20K\ast \ast ;theNEARuniverse-U\ast \ast median$19K.
• 100% of universe-U buckets have ≥ $1,000 volume.
• This is the cheap-bucket cohort specifically — the v1 thin-tail caveat (~$1k–$9k on deep longshots) does not gate this trade. The blocker is sample size / CI, not liquidity.

The progression, and why it is a CANDIDATE not a GO

Progression — both levers help and stack

buy-blind ≈ −15% → near-forecast (select) +1.2% → near-forecast + 18:00 exit +9.6%. Forecast-proximity and exit-timing are independent levers that each add edge and compound.

CANDIDATE, not a confirmed GO

Even the best version is +9.6% mean with an 80% LCB of −6.9% on only n=125 over a single month. The lower bound is still negative — this is a promising candidate, not a statistically confirmed strategy. Do not size real capital on one month of one city-set.

This is direct evidence for the 2026-06-11 “intraday nowcast / information-timing” pivot: the market reprices toward the realized high through the afternoon, so the edge is picking the right cheap bucket via forecast (and ultimately a live nowcast that tracks realized temperature), not the drift itself.

Next steps (v3)

1. Extend the data window to several months (and ideally more cities) to tighten the CI — the single binding constraint is sample size, not signal or liquidity.
2. Tighten + sweep, free off cache: restrict to dist ∈ {0, +1} (the winning buckets) and sweep the exit hour 16–20h to find the convergence sweet spot.
3. Add a live intraday nowcast (realized-temp tracking) to sharpen winner-prediction beyond the static forecast — the natural bridge from “forecast-conditioned” toward the information-timing edge the pivot targets.

Charts and the full report live at research/report.md + research/charts/.

Caveats

Read before trusting the numbers

• The v2 candidate rests on n=125 over a single month (4 cities). The 80% LCB is still −6.9% — promising, not confirmed. Treat the +9.6% as a hypothesis to validate on a wider window, not a tradeable edge.
• Forecast = Open-Meteo previous-runs proxy. Literal historical Wunderground forecasts are NOT retrievable (verified) — the lead-0/1/2 figures are a reconstructed stand-in, not the forecasts a trader would have literally seen. The v2 forecast-proximity filter inherits this proxy risk: a live trader’s actual forecast may bucket differently.
• Realized high is best-effort via TWC — the same authoritative-source posture as fix-resolve-twc-primary-source.
• Thin-bucket fills make longshot paper returns optimistic — per-bucket liquidity is ~$1k–$9k and tails are wide-spread; real fills on longshots would be worse than the paper marks.
• Polymarket has 0 trading fee, so modeled cost = spread only. Slippage beyond top-of-book is not modeled.

Related

• wx-research-crate — the wx-research crate (branch research/intraday-drift) this study runs in
• wx-market-prices-history — CLOB /prices-history client supplying the intraday curves
• wx-market-trades-client — data-api /trades client
• polymarket-weather-market-structure — market structure, bucket layout, ICAO/whole-°C resolution, liquidity
• fix-resolve-twc-primary-source — TWC as the authoritative realized-high / resolution source
• weather-bet — project overview