Ricefan

YC S26 EdotEnv: self improving RL benchmarks from Quant data | Quant @ G-Research

[Ricefan @RRicefan]

Fun observation 3/3:
The portfolios are quite concentrated in small amount of coins. Clearly the risk modelling tools were not utilised well.

[Ricefan @RRicefan]

GPT5.5 made the most profit in Crypto 2021, but Opus48 had the highest Sharpe.

We gave AI agents proper Quant Trading tools to operate in the market, with agents calibrating tools themselves. (They are not trading by tea leaf reading!)

Here's what the individual models did:

[Ricefan @RRicefan]

Fun observation 2/3:
Grok is a bag holder! Analysing his trades during his drawdowns shows that in most of them no trades happen.

[Ricefan @RRicefan]

@data4rtist @annanay 2. the models have forecasting tools, so we’re evaluating when/how it uses tools to make decisions - not whether it can trade by reading tea leaves from pretraining memory.

[Ricefan @RRicefan]

@data4rtist @annanay Yeah leakage is a real concern. My 2 takes:
1. In this setup the models only see past numerical data, no metadata for context. Empirically we find that models are not able to produce good forecasts like this.

[Ricefan @RRicefan]

@TradfiBaby Great idea!

[Ricefan @RRicefan]

@lustcouncil DMs open now!

[Ricefan @RRicefan]

@acrizzle Try your favourite LLM on past timeseries and show me the data leakage

[Ricefan @RRicefan]

Blue is btw the same setup calibrated by a Citadel Trader.

[Ricefan @RRicefan]

@PariLatawa thanks, and you! Will see you in the batch :)

[Ricefan @RRicefan]

Three YC rejections later, I finally got into YC S26. Looking back, this was probably the right time.

Models are improving faster than static benchmarks can keep up. We're building RL evaluation environments from Quant Trading data: benchmarks that adapt as models get better, so there's always a harder hill to climb.

[Ricefan @RRicefan]

@TheMaxTP tell him to watch out! :P And let me know if he wants to chat.

[Ricefan @RRicefan]

@kwertytwo @annanay @RJCcapital @DarkPoolKnight 💯

[Ricefan @RRicefan]

Def don’t miss out

[ak0 (semis/acc) @annanay]

Hosting a X space with @RJCcapital and my kwants including @RRicefan, @DarkPoolKnight and others in 3 hours! (midday eastern). We'll be talking about:
- the state of quant
- leaving quant to found in AI
- are semis the new memecoins

And more!

Comment below any questions you'd like to ask us

[Ricefan @RRicefan]

Three YC rejections later, I finally got into YC S26. Looking back, this was probably the right time.

Models are improving faster than static benchmarks can keep up. We're building RL evaluation environments from Quant Trading data: benchmarks that adapt as models get better, so there's always a harder hill to climb.

[Ricefan @RRicefan]

@SamuelAlbanie Me!

[Ricefan @RRicefan]

@JakeFleshner Self-improving benchmark