A brainstorming idea

Can an LLM learn to trade on its own?

The model makes trading decisions, while coding agents improve the system around it.
30 September 2026 · Emil Valeev for Manaraga

Where the idea came from

~2022

A Go bot for dYdX. It turned out I needed to be a trader.

order book slippage funding liquidation margin dropped it
2026 · PARTYstation

Tests capture the old implementation; coding agents generate a replacement until the tests pass.

acceptance ≈85%
Now

With a good harness, LLMs can run autonomously, and self-improvement loops have emerged. Returns make an ideal metric.

Two ideas →
Why not 3 years ago? Back then, I had to be the trader and write the code myself.
1 / 8

Two ideas, two independent choices

who decides?who improves? Trading modelIDEA 1 code human coding agentsIDEA 2 human coding agentsIDEA 2 ● now: the model trades both ideas together classic algo trading · 2022 ✕ coding agents write strategies
Idea 1 hypothesis: a capable model makes good decisions if given the right inputs with the noise removed.
2 / 8
Public arenas and papers · October 2025 – September 2026

Can models trade? Nobody knows yet

43
trades by the winner. About 860 are needed to distinguish skill from luck.
Oct–Nov 2025
0 of 14
models beat BTC during a rising market.
Aug–Sep 2026
43%
of the winner’s profit went to fees.
Oct–Nov 2025
−13% → +19%
same model, different scaffolding.
2026 paper
In the only paired comparison, scaffolding changed the result more than the model did.
Sources ↗ Alpha Arena · TradeRank · paired scaffolding comparison
3 / 8

The idea: a gate measures returns; a coding agent explores variants

Coding agent writes variants variant 1 variant 2 variant 3 variant 4 variant 5 variant 6 trading system variants GATE ✕ lower PnL ✓ higher PnL ✕ lower PnL ✕ lower PnL ✓ higher PnL ✕ lower PnL The best survive seed for the next generation next generation
The gate runs a variant on market data. The one with higher PnL survives.
4 / 8

Divide and conquer: measure each component separately

One system one metric: PnL split into components → runtime model Proposes trades can split further combinatorial explosion the right approach is unclear runtime model Reviews trades tune with evals dataset known good / bad answer = expected? metric coding agent modifies the reviewer
Once tuned, a component leaves the search space. It all comes down to the dataset.
5 / 8

Drawbacks: evaluation cost is the main constraint

1. Every model run costs money
hourly runtime · ~50%
experiments · ~50%
Claude Max 5x subscription
LLM configurations
a handful
Python strategies
hundreds, almost free
But judging code quality still takes expertise.
2. The model remembers the past
current model training data — remembered clean data new model model release
Evaluate only on data after the model’s release. A new model means collecting fresh data.
Where I am now: at first the model made no trades; now it trades at a loss.
6 / 8

What the company already has

gate
=
evalsPeople who build them.
a system accepted by tests
=
PARTYstationWe built a system where tests are the main gate, rather than a person.
LLM + crypto
=
Maneki GardenAI for DeFi portfolios, an adjacent domain.
7 / 8

If it works, it could become a product

Ideally, the investor earns returns without being a trader.

Pain point
Passive incomeMoney that works without its owner.
People already pay for workarounds
Copy trading, bots, AI agentsExchanges are releasing tools for AI agents in 2026.
Revenue model
Subscription or % of profits
pain pointPassive income
why nowWhere the idea came from
what we haveWhat the company already has
8 / 8