wxedge — Temperature Market Edge

Overview

Predicting the resolved bucket of a Polymarket daily-temperature market is an intraday problem, not a forecasting problem. The daily max happens between 13:00–17:00 station-local time, so at/after 15:00 the running max is almost locked while the market still prices the winning bin cheaply (avg ~0.13 while it resolves ~79% of the time). Everything below is measured against real market settlement, not model output.

Project: /home/user/wxedge/ · full rules: STRATEGY.md · evidence: reports/*.md

When to Use

• "Which bucket will resolve for on ?" / "ควรซื้อ bin ไหน"
• Anything about hit rate, cheap bins (<$0.25), intraday (14:00–18:00 local) predictions
• Validating or comparing a temperature-market strategy against alternatives

Do NOT use for: general weather questions, other market types, or any day-ahead-only bet (day-ahead top-1 is only ~50%).

The Three Facts (check before any prediction)

1. Resolution is station-specific, local-calendar-day, whole-degree. Each market names its station (nyc=KLGA, london=EGLC, hong kong=HKO, moscow=UUWW, taipei=RCSS …). Never use city-center coordinates, never use a different day boundary, never re-convert °F/°C (US bins are whole °F; Wunderground units=e is already whole °F).
2. Observations must be IEM hourly METAR (mesonet.agron.iastate.edu, data=tmpc, report_type=3, tz=<city tz>) — it matches the resolving source 94%. Do NOT use NWS 5-minute observations (api.weather.gov): only 56% match, runs hot ~+0.3°F. IEM end date is exclusive → request day+1.
3. Timing decides everything. Predict at/after 15:00 station-local, best at 16:00. Local-time errors are the single most common way to lose money here.

Workflow

1. Check the station's local time now — it must be 15:00–17:00 (a late-peak city: 17:00–18:00; see references/playbook.md). If it is earlier, say so and wait; do not bet.
2. Run the tool (never hand-compute probabilities):

   python3 /home/user/wxedge/cheap_live.py                  # scans cities currently 14:00–19:00 local, default rule p>=0.90
   python3 /home/user/wxedge/cheap_live.py --cities denver,seattle --pmin 0.35 --json
   python3 /home/user/wxedge/cheap_live.py --log            # log recommendations for forward test

   For analysis/backtest instead of today's bet: intraday.py --days 120 --hours 14,15,16,17 --cities all.
3. Apply the gates (below) to each candidate. If nothing passes, the correct answer is "no bet today" — opportunities are ~4/day at 16:00 across all cities, not every city every day.
4. Report with the evidence line: state the rule used, the model probability, the ask, and that ROI figures are gross (last-trade prices, no fee/spread). Never claim a result you did not just compute in this session.

Gates (all must hold)

Gate	Rule	Why (measured)
Time	15:00–17:00 station-local (16:00 best)	14:00 = 61.4% hit, 16:00 = 85.3%, 17:00 = 88.5%
Price	ask 0.02–0.25	below 0.02 wins only 6.0% (n=182) — the market sees finer 5-min data than we do
Model	p ≥ 0.90 (wide base) · p ≥ 0.97 (sharp) · p ≥ 0.35 (max volume)	16:00: p≥0.90 → 78.7% @ 0.134 (+488%); p≥0.97 → 90.9% @ 0.136 (+570%, n=99); p≥0.35 → 70.5% (+422%)
Size	1–2% of capital per bet, cap ~10% total	70–79% win rate still loses runs of bets

Never Do (each was measured and failed)

1. Never buy bins priced under 0.02 — 6.0% win rate.
2. Never predict before 14:00 (~20–30% win) and never bet day-ahead without p ≥ 0.65 (top-1 is only 50.9%).
3. Never use take-profit (+30/+55/+100%) or a forced 17:00 sale — both cut ROI. Hold to resolution; an optional −25% stop-loss is the only exit that helped.
4. Never use NWS 5-minute observations (56% match) — IEM METAR only.
5. Never port a repo's σ formula as if it were the edge — swapping σ formulas changes results ≤0.1 pt; the real drivers are (a) removing per-model bias before blending (+17.4 pts) and (b) using post-peak observations (+52 pts).
6. Never select cities using past results (train-picked subset: 80% → 43% out-of-sample).
7. Never use "adjacent to the favourite bin" as a rule — negative ROI at 16:00–17:00.
8. Never quote ROI without saying gross: hourly last-trade prices, no fee/spread/slippage, 85 days of a single season.

Quick Reference (headline numbers, test period, real settlement)

Question	Answer
Day-ahead top-1 / top-2	50.9% / 80.2% (n=1,892, Brier 0.608)
15:00 / 16:00 / 17:00 top-1	76.0% / 85.3% / 88.5% (n=4,010)
Cheapest profitable rule	ask 0.02–0.25 & p≥0.90 @16:00 → 78.7% win @ 0.134 → +488% ROI (gross)
Best win-rate rule (price ≤ 0.30)	ask 0.02–0.25 & p≥0.97 @16:00 → 90.9% win @ 0.136 → +570% (n=99, test 90.5%)
Raise price cap to 0.30?	No — widening 0.25→0.30 lowers ROI (488%→371%): pay more, win the same
Fee/spread stress	with 2% fee on winnings and paying 2 ticks more: p≥0.90 → +401%, p≥0.97 → +471%
vs. alternatives	ours 88.8% intraday > PolyWeather+hard-floor 77.3% > weatherbot 45.5% day-ahead (own paper −22%) > polyBot 34.7%
Best single borrowed idea	hard floor: any bin below (max_so_far − 0.5°C) is impossible (+39.7 pts)

More tables (per-city timing groups, ROI by rule, traps): references/playbook.md.

Common Mistakes

• Treating a model's max-probability bin as the answer without checking price → wins 77% but ROI only +24% (PolyWeather's pattern).
• Using the city's timezone guess instead of the station's actual local time.
• Reading prices-history?interval=max (empty for markets older than ~1 month) instead of startTs/endTs.
• Reporting a number from a stale report file instead of re-running the command.

Evidence

All figures above come from reports/headtohead.md (head-to-head of 4 strategies on identical inputs), reports/cheap_bets.md (220,338 bin×day×hour prices), reports/hitrate70.md (4,010 city-days vs real settlement), reports/superpowers_vs_strategy.md (methodology comparison). Reproduce with python3 headtohead.py --split 2026-08-15.