Hedge funds have research desks. Now you do.

Name: TailFifty public research record
Creator: TailFifty
License: https://creativecommons.org/licenses/by/4.0/

Type a question about oil, corn, Bitcoin or the weather. Twelve AI analysts pull the data, run the tests, and fight it out. You get a probability, the evidence, and a scorecard the calendar fills in.

An agentic investment research platform, not a stock picker. An experiment that shows every step and keeps score on itself.

What your desk works from

Prices, positioning, degree days, storm tracks, funding rates, injury reports. 156 series, refreshed every night, each stamped with the moment the desk could first see the number.

30 NOAA forecasts
30 NOAA observed
28 weather stations
28 prices
15 CFTC positioning
9 crypto spot
6 Kalshi
4 crypto derivatives
3 injuries
2 scores
1 hurricanes

290,886dated observations
151forecasts with a resolve date
972links from claim to evidence

This season: NFL, scored against the line

Three games a week. The committee's number, the Kalshi price at the time, the final score, a Brier for each. See every game

What happens after you ask

1
You ask.
A yes-or-no question with a date. Will corn rise 8% by harvest? Will Bitcoin drop 10% in two weeks?
2
Twelve AI analysts dig.
Each seat pulls the series and picks the tests. Code runs every test and hands back the numbers.
3
Each seat writes a sealed view.
A probability, the evidence, the strongest case against, and what would prove the view wrong.
4
The seats argue.
Names hidden. Every challenge answered. Every change of mind pinned to one piece of evidence.
5
The portfolio manager calls the number.
Seats with a better track record weigh more. You get one probability and the whole trail.
6
The calendar grades the call.
On the date, the store scores the forecast against what happened, and against the market where one priced the same question.

House rules

Code enforces every rule below. No prompt can bend a rule.

No analyst ever invents a number.: Code runs every test. An analyst sees only the results. agents/runtime.py
No forecast without a date and a rule.: The store scores every forecast on the resolve date. Misses stay on the record. research/calibration.py
No peeking at the future.: Every read takes an as-of time, the moment a number went public. A backtest sees only what was public that day. data/store.py
No groupthink.: Twelve seats, twelve goals. Views stay sealed until the argument. The skeptic always sits. debate/engine.py
Track record sets the vote.: A seat that calls more right weighs more. A seat that misses weighs less. AgentPerformance
Findings stack up night after night.: A miner tests every signal against every target while you sleep. The desk cites the library instead of redoing the work. research/miner.py
No new angle, no seat.: A seat with nothing to add spends nothing, so your question gets the analysts who matter. agents/gating.py
Risk is arithmetic.: The desk proposes a trade on paper. Limits size the trade or refuse the trade. No broker is connected. risk/engine.py

Your twelve seats

Each seat is an AI model with a role prompt, not a person. Calibration: does a 30% call come true 30% of the time? Read every debate.

Commodities Analyst physical balances, storage, the curve.
Weather Analyst degree days and forecast revisions.
Public Markets Analyst sector and factor exposure.
Crypto Analyst funding, open interest, leverage.
Sports Analyst injuries and rest against the line.
Quantitative Researcher regressions, event studies, out-of-sample tests.
Historical Analog Analyst base rates across comparable windows.
Market Positioning Analyst what is already priced in.
Causal Skeptic confounders, leakage, placebo tests.
Statistical Reviewer seventeen questions, no mercy.
Risk Officer assumes you are right, lists how you lose.
Portfolio Manager weights by calibration. Never averages.

Brier score. Lower wins.coin flip = 0.25

Sports0.007 n=1

Crypto0.040 n=1

Markets0.082 n=2

Quant0.199 n=5

Causal Skeptic0.201 n=5

Committee0.201 n=5

Positioning0.213 n=5

Commodities0.310 n=3

Weather0.320 n=3

Analogs0.321 n=3

Calibration: the desk keeps score on itself

Where TailFifty started

151 won 93 lost 14 open or void

A $50 Kalshi account. 258 bets, all public, posted for the record and not as picks. A 62% hit rate that ran seven cents worse than the closing line, the market's final price: luck, not skill. The tools that proved the luck became the desk. The ledger stays up.

The bar the desk must clear

Beat the market's Brier score on 60 resolved forecasts that a market also priced. Until then, the honest word is "not yet".

5 of 60 resolved0 priced by a market

Committee Brier 0.201. See the record. Losses stay in. Past forecast accuracy does not predict future accuracy.

Get a seat

Your questions. Your desk. Every answer with the trail and the score. Seats open in order. No price yet.

Built by

Harrison Johnson built TailFifty to get a research desk worth arguing with. Models pick questions and weigh arguments. Code does the math and keeps score.

TailFifty runs every night under a spend cap and records every model call, so the history replays for free. Harrison wrote the code with Claude Code and made the judgment calls.