Fifteen seasons, 300 rounds, five opinions

The five underwriters disagree about how much the distant past tells you about next season. That cannot be settled by argument, only by capital. This runs real Nigerian rainfall from 2011-2025 through the same auction rules the contract enforces and reports who was still solvent at the end.

Reproduce it yourself: pnpm sim. No network, no keys and no LLM calls — 20 buyers, 5 underwriters, 8 sites, seed 20261112, against a rainfall fixture pulled 2026-11-12. Same seed and same fixture give byte-identical output, and a test asserts that.
Rounds run
300
2 found no bid at all
Policies written
298
each fully collateralised by its underwriter
Paid out
111
37% of policies written
Best and worst
Recent
Longrun lost 1,896.51 USDC of its own capital

Where the capital went

AgentWonBidDeclinedPremiumsClaimsLoss ratioFinal capitalProfit
uw-1 Longrun
Every season since 1981 counts equally; the climate is noise around a mean.
241267333,033.494,930.00163%103.49−1,896.51
uw-2 Recent
Only the last 15 seasons tell you anything; the old ones are a different climate.
353000326.15160.0049%2,166.15166.15
uw-3 Blend
Split the difference and charge a little more for not knowing which is right.
2300043.25130.00301%1,913.25−86.75
uw-4 Cautious
Price the worse of the two views. Survive first, compete second.
030000.000.00n/a2,000.000.00
uw-5 Capacity
Blend the views, then ration the last of the capital by charging more for it.
203000369.46420.00114%1,949.46−50.54

Loss ratio is claims paid over premiums earned. Above 100% means the agent paid out more than it was paid to take the risk. Every agent started with 2,000.00 USDC.

What actually happened

Longrun prices on all 45 seasons since 1981, so it quotes the lowest premium of the five, wins most of the market — and pays for it. Recent, which trusts only the last fifteen seasons, is the only agent that made money. Cautious wins nothing at all and ends with its capital intact, which is exactly what “survive first, compete second” buys you.

Nobody told these agents the climate is drying. The rainfall did, and the ones that would not listen lost their own capital. That is the whole mechanism: a market where mispricing is punished with the underwriter’s money rather than the farmer’s.

Capacity’s behaviour is emergent rather than scripted. Its load rises as its book fills, so it stops being the cheapest bid late in a season and Recent picks up the business it drops.

Clearing price by season
Mean winning premium as a share of the payout, across every round that cleared.
0%10%20%30%23.2%20112013201520172019202120232025

Every round

All 300 rounds with every bid, including the declines, are in sim/results/results.json. The most recent twenty are below.

#SeasonSiteStrikePayoutWinnerCleared atRainOutcome
3002025makurdi120.680.00no bid—109.7—
2992025ibadan104.750.00recent7.06129.4expired
2982025nsukka71.750.00blend19.6640.9paid out
2972025enugu71.130.00capacity10.2653.6paid out
2962025maiduguri62.280.00recent16.9441.8paid out
2952025sokoto92.050.00recent7.06257.9expired
2942025kano63.250.00capacity12.6642.7paid out
2932025kaduna89.930.00longrun7.04144.9expired
2922025makurdi120.680.00capacity31.86109.7paid out
2912025ibadan97.750.00recent3.53129.4expired
2902025nsukka60.550.00capacity14.2140.9paid out
2892025enugu70.430.00longrun6.2653.6paid out
2882025maiduguri60.080.00recent11.2941.8paid out
2872025sokoto80.750.00recent7.06257.9expired
2862025kano66.750.00longrun11.7442.7paid out
2852025kaduna89.930.00longrun7.04144.9expired
2842025makurdi118.480.00capacity27.38109.7paid out
2832025ibadan104.750.00recent7.06129.4expired
2822025nsukka60.550.00longrun10.4340.9paid out
2812025enugu83.130.00longrun8.6153.6paid out
Buyers pick a strike between the 15th and 25th percentile of their site’s history before that season, and will pay at most 40% of the payout. Agents price only on seasons that ended before the one they are quoting — a backtest that peeks looks brilliant and tells you nothing.