wind_anchor

1. There is no wind market. Not on Kalshi, and essentially not on Polymarket.

I pulled Kalshi's full series catalogue — 12,940 series across eighteen categories — and searched every ticker and title for wind, gust, gale and derecho. Zero matches. The exchange lists 334 weather series and they cover daily highs and lows in dozens of cities worldwide, rain, snowfall, hurricanes by landfall point, air quality, earthquakes, volcanic eruptions, drought level, Lake Mead, arctic sea ice and cloud cover during an eclipse. Wind is the conspicuous absence. Our own ingest agrees: 0 of 539,344 stored Kalshi rows.

Polymarket has listed exactly three, out of 124,745 distinct titles we hold: “Will the highest recorded wind on Mt. Washington in July be at least 90 / 95 / 100 mph?”, carrying $17,081, $28,458 and $18,395 of volume, all closed 31 July 2026 and all resolved YES. Mount Washington is a single 6,288-foot summit with a cult reputation and one official observer — the opposite of a sixteen-city blend, and 180 km from the nearest metro we poll. Polymarket filed all three under other, and the starter bot we ship filters on categories: ['weather'], so even the one wind market that has ever existed would have been invisible to it.

That absence is probably rational rather than an oversight. A gust is the noisiest settleable weather variable there is — it depends on the exposure of one anemometer and on whether you average over three seconds or five — and “will Denver gust 45” has none of the cultural pull of “will it hit 100 degrees”. This is the market declining, not the market not having noticed.

2. Checked against the wind that actually blew, the gate does not sort windy days from calm ones.

Gusts can be verified, which makes this signal unusual in this dictionary. The National Weather Service publishes the peak gust recorded each day at every major airport, free, through Iowa State's Environmental Mesonet. So for all 241 stored readings I looked up what the wind actually did at all sixteen cities across the two days each reading was forecasting, and applied the identical “largest of sixteen” rule to the observations that the poller applies to the forecast — so both sides of the comparison are selected the same way.

Gate open: a gust above 45 mph was actually observed somewhere on 51.9% of polls (14 of 27). Gate shut: 63.1% (130 of 206). The base rate across everything was 61.8%. Two-proportion z = −1.13, p = 0.26, and the 95% interval on the difference runs from −31.2 to +8.7 percentage points. Grouping by day instead of by poll, to be conservative about how much adjacent readings repeat themselves, gives 71.4% on the 14 open days against 87.5% on the 16 shut ones — same direction, same lack of significance. Ranked as a score rather than a threshold, the forecast's ability to order days by what the wind did gives an AUC of 0.513 at 45 mph and 0.512 at 40 mph: a coin flip. It does better at 50 mph (0.631), which is the one hint in here that something real is buried in the series.

So the honest verdict is not “the signal is backwards”. It is that there is no evidence it sorts windy from calm, and the point estimate leans the wrong way. With 27 open polls across 14 days this test cannot rule out a modest edge in either direction — but it can rule out the strong claim the product makes.

3. The number we threshold runs ten miles an hour below the reality it is named after — and it is the maximum, not the forecast, that breaks.

Over the same thirty days, the observed two-day national peak gust averaged 47.9 mph. Our stored forecast averaged 37.9. Mean error −10.0 mph, mean absolute error 11.6, and the forecast was the higher of the two on only 15% of polls.

The obvious explanation is wrong, and I checked. The poller reads city-centre coordinates while the observations come from airports — up to 30 km away at Denver. So I re-ran the same Open-Meteo product over the same thirty days at both coordinate sets. The difference across all sixteen metros was −0.01 mph, and on the national maximum −0.06 mph. Location explains none of it.

The real explanation is arithmetic. At the one city the forecast names as windiest, it is essentially unbiased — 37.9 forecast against 37.8 observed, median error +1.0 mph, higher on 51.9% of polls. The per-city forecast is good. But a maximum over sixteen smooth model forecasts is systematically smaller than a maximum over sixteen noisy real observations, because reality gets sixteen independent chances to spike and a smoothed field does not. The aggregation, not the meteorology, is what puts ten miles an hour between the gate and the world. Every threshold anyone sets on this series is therefore being set in units that are not the units on the label.

4. A bot called “High-Wind Watch” fires below the weakest official wind product in America.

The National Weather Service publishes its criteria, and I read them rather than recalling them. A Wind Advisory — the mildest wind product there is — is issued for sustained winds of 31 to 39 mph for an hour or more and/or wind gusts of 46 to 57 mph for any duration. A High Wind Warning requires sustained winds of 40 mph or more for an hour, or gusts of 58 mph or greater. A High Wind Watch uses the same numbers for a possible event. An Extreme Wind Warning starts at 115 mph.

Our shipped threshold is 45 mph — one mile per hour below the bottom of the advisory range, and 13 below the line that means damage. The bot's name is “High-Wind Watch”; the National Weather Service's High Wind Watch begins 13 mph further up. And 58 mph was crossed exactly once in 241 readings (a 62 on 11 August). Local offices adjust these numbers — Denver's foothills use a higher bar because 45 mph there is a Tuesday — but adjusting them upward only makes the point worse.

Read plainly, {peak, above:45} does not mean “a windstorm is coming”. It means “a model currently believes that at one of sixteen points, on one of the next two days, the peak gust will be a bit breezier than usual”.

5. Thirty days of memory, taken in the calmest month of the year.

Wind is the most seasonal variable in this dictionary, and our thirty-day retention window sits exactly on the annual floor. Rebuilding the same statistic from the ERA5 reanalysis across 3,875 two-day windows from 2016 to 2026, the shipped 45 mph setting would have been open on 2.1% of days in July and 4.1% in August — against 40.5% in March, 32% in January and February, and 31% in December. Our exact calendar slice, across ten years, comes out at 3.7%.

This changes one verdict for the better. The docstring's {metric:'mean', above:30} example never opens in our data — the highest mean_gust we have ever stored is 29.5. It would be easy to call that a dead setting. It isn't: over the full reanalysis the national average clears 30 on 16.1% of days and has reached 44.5. It is a winter setting pointed at a series we only ever look at in summer. Our memory is the defect, not the number.

It also means everything else on this page — every duty cycle, the verification, the ten-mile-an-hour gap — was measured at the annual minimum, on the kind of wind that comes out of a summer thunderstorm rather than off a winter cyclone. Those are different physics, and a signal can be useless at one and useful at the other. We delete the evidence that would settle it every thirty days.

6. Most of the movement is the winner changing, not the weather changing.

The stored number is a maximum, so it jumps whenever a different city takes the lead. Decomposing the variance of all 241 readings: the hour of day explains 3.0% — so this signal does not inherit ensemble_anchor's clock defect, which is now confirmed for the third weather signal running. The calendar day explains 27.0%. But the identity of whichever metro happened to be windiest explains 29.0% — more than the date does.

The split makes it concrete. On the three days when one city stayed windiest from midnight to midnight, the reading moved by a median of 3.0 mph. On the 28 days when the winner changed hands, it moved by 14.5 mph. There were a mean of 3.68 different winners per day, and the gate changed state 38 times in 241 polls at the 45 mph setting (54 times at 40). Denver won 50 times and Chicago 46; Los Angeles and Seattle never won at all.

And the stored label is usually wrong: the metro the forecast named as windiest really was the windiest on 18.0% of polls. That is nearly three times chance (6.25%), so the forecast knows something — but four times in five, windiest_metro names the wrong city.

7. The backtester silently ignores it — and here the standard excuse is wrong in letter but right in substance.

wind appears nowhere in quants/backtest.py: not in the supported list, not even in the documentary _UNSUPPORTED_ENTRY comment. It is caught only by the catch-all that rejects any key it doesn't recognise. The consequence is the one that matters: a backtest of a bot with this gate runs without the gate and still returns an equity curve, flagging unsupported_rules. You get a number that looks like a test of your idea and isn't.

The poller's docstring justifies this by saying “Open-Meteo's free forecast is forward-only (no archive here), so the signal is live-only”. In letter that is false — I verified both the ERA5 archive and the historical-forecast archive return HTTP 200 for wind_gusts_10m_max, and pulled 3,876 days from one of them for this page. But backfilling from either would be actively misleading, and I can put a number on it: over our own thirty days the near-analysis archive produces a national peak averaging 33.5 mph that clears 45 on 0.0% of days, where the live feed produced 38.0 mph and 11.6%. A backtest built from the archive would conclude the gate essentially never opens. The excuse is wrong about the archive existing and right about not trusting it.

8. The “level up” the starter bot suggests is a one-in-550-day event.

The starter's blurb offers: “flip it to ‘below’ to trade only dead-calm stretches.” The docstring's own example of that is {metric:'peak', below:25}. Across our entire record the lowest reading is 27 mph, so it has never opened; across ten and a half years of reanalysis it would have been true on 0.18% of days — about one day in 550. Sixteen large American cities are essentially never all calm at once, which is a property of taking a maximum over sixteen cities rather than a property of the weather. Nothing in the builder warns you.

9. The usual validator gaps, and the usual off-by-one prune.

{metric:'peak', above:45, below:25} validates cleanly and renders as “only when windiest-metro peak wind gust > 45 mph and windiest-metro peak wind gust < 25 mph” — two opposite regimes in one sentence, unsatisfiable, no warning. {above:1000} validates. {below:0} validates and can never open, since a gust is never negative and the comparison is strict. {metric:'mean', above:0} is always open. In fairness the units here are unambiguous — the field says mph and the stored value is mph, which is more than finance_anchor can say.

And the prune keeps 726 hours, not the 720 it intends, because it compares an ISO timestamp containing a T against SQLite's datetime() format containing a space, as text. This is the ninth poller in this dictionary with that exact bug.