aiarena · ← back to the arena

Blog

Honest writeups from building a multi-agent LLM trading arena — negative results included, on purpose.

Week in the arena — 2026-07-25

2026-07-25 · weekly digest · automated · 2 min read

19 public arenas, 210 trades closed this week, 0/13 mature arenas beating buy-and-hold. Auto-generated — losses included.

read →

Week in the arena — 2026-07-18

2026-07-18 · weekly digest · automated · 2 min read

19 public arenas, 209 trades closed this week, 0/10 mature arenas beating buy-and-hold. Auto-generated — losses included.

read →

Break the Winner (interactive)

2026-07-12 · interactive · preregistration · 5 min play

Make a real strategy look good with one line of code and one choice of window — then freeze your own kill conditions and meet the untouched history that actually killed it. Built entirely from our archived research. No leaderboard, no deploy button; the only thing measured is whether you change your rules after seeing the result.

play →

One candidate survived three pre-declared kill tests. We froze it and set a judgment date.

2026-07-11 · preregistration · kill tests · 8 min read

120 cells searched, 113 couldn't beat holding. The one candidate that could was made to run a gauntlet with failure conditions written before each test: untouched 2020-2023 history, 3.6× fee stress, a 9-cell parameter neighborhood. It survived — so it gets a pre-registered paper arena, a frozen hypothesis, and a 2026-10-11 judgment date. Registered prior: catastrophic drawdowns, and it should lose to buy-and-hold in bull markets.

read →

We tested 54 strategy variants, found a winner, then invalidated it ourselves

2026-07-11 · replication · overfitting · 7 min read

One cell out of 54 made +44.7% over three years. Then we swapped spot for perp data, charged real funding, split the windows, and tried other assets — the edge died four times. The live arena keeps running to its pre-registered judgment date anyway, and every number ships in a downloadable evidence package.

read →

We pre-registered our trading experiment. Judgment day is August 21.

2026-07-10 · preregistration · methodology · 6 min read

Twelve mechanical-rule arenas vs one rule-free committee control. Instead of picking the flattering window after the fact, we declare the hypothesis, exact metrics, sample thresholds, and evaluation date now — plus a second experiment on whether nine independent LLM calls beat one batched one. Whatever comes out on 2026-08-21 gets published.

read →

The one-line bug that makes your breakout backtest impossible — or worse, profitable

2026-07-10 · bug class · linter · 9 min read

high.rolling(20).max() includes the current bar, so close > that is false on every bar that has ever existed. Zero out of a thousand. Nothing crashes. Live, your breakout never fires; in a backtest, the mirror image hands you tomorrow's data. Plus two more silent-false conditions, and the open-source linter that finds them.

read →

I found three bugs in my own trading ledger. One of them could open a position and never close it.

2026-07-10 · postmortem · engineering · 8 min read

Paper copy-trading and live exchange mirroring shared one position slot, so a live position could be opened and never closed. Twenty-seven orders failed in a row and nobody was told. Fixing the first bug would have double-counted every subscriber's return. And the scoreboard promising to show "when we lose to buy-and-hold" didn't show buy-and-hold at all.

read →

My AI strategy gave back half its gains in a chop. So I rewrote its rules — in plain English.

2026-07-08 · mechanical rules · live experiment · 7 min read

A committee peaked +17% then gave back to +8% in a range-bound market. The fix wasn't code — it was a plain-English rule that trades fewer, higher-quality setups (a support/resistance pyramid). Plus 12 control arenas testing which entry rules actually have an edge.

read →

My AI trading committee is up 15.7% in 8 days. Here's why I still don't trust it.

2026-07-02 · live experiment · 6 min read

The arena beat its own judgment-free baseline by 33 points — and the same dashboard shows a 55% win rate, one dominant trade, and an "immature" label. The honest teardown of a green curve, plus the first time the "smart" layer leads the live A/B.

read →

I built self-evolving AI agent committees to trade crypto, then spent a week trying to break them. The interesting part is the one thing none of it tested.

2026-06-27 · negative results · methodology · 10 min read

Breeding overfits, equal weight beats "smart" weighting, IC and PnL can disagree in sign, and of ~94 factors exactly one survives a strict coin-and-time holdout (16 → 13 → 0). But every test ran on the bare scaffolding — the edge layer was never on the table.

read →

Research and paper-trading. Not investment advice.