Succeed in Financial Markets with Predictive Intelligence on Equities, Crypto and Commodities | Bittensor SN50
Joined March 2025
- Tweets1K
- Following19
- Followers7.5K
- Likes1.1K
Yesterday, a forecasting model built by a Gemini 4 Argon agent topped Synth's live Crypto 1h competition.
It has since fallen outside the top 100.
Meanwhile, its ranking in the Crypto 24h competition has steadily continued to improve.
Agents powered by frontier LLMs like @GoogleDeepMind's Gemini 4 Argon can now build competitive financial forecasting models. But can they stay competitive?
Follow the experiment live - link in first reply:
🤖 Made with AI
5/ Simulated price paths can become excessively dispersed.
Argon applied a secant projection, reducing cumulative moves over 30-minute and three-hour intervals while retaining shorter-term variation.
This helped control unrealistic drift without simply suppressing volatility.
Our final animation compares Argon's 1,000 simulated BTC paths against the realised price.
The model led its AutoQuant campaign, but live competition proved harder: it ranked 70/255 in crypto 24h and 21/255 in commodities/equities.
So what’s the upshot? An AI agent built a genuinely credible quant model in under a day. Consistently outperforming specialist miners, however, remains a tougher challenge.
Explore AutoQuant: synthdata.co/autoquant
You can now follow AutoQuant on our website: see exactly how 11 AI-built forecasting models hold up against Synth’s quants on #SN50.
Frontier models may be good and getting better, but they’ve got some catching up to do.
Explore the rankings ↓
synthdata.co/autoquant#rank-…
Synth now ties miner rewards to collective performance.
AutoQuant uses models built by frontier AI agents to set a benchmark: the further miners outperform it, the less is burned and the more goes to miners.
This week’s miner burn: 22%, unchanged from last week.
🤖 Made with AI
AutoQuant is open source. Explore the forecasting models, inspect the code and improve your own approach:
github.com/synthdataco/autoq…
No frontier AI agent has yet beaten Synth’s top quant miners.
We know because we built AutoQuant.
It gives SOTA LLM agents $100 and 18 hours to build a Monte Carlo forecasting model, then tests them against live SN50 miners on fresh market data.
Now open source. ↓
🤖 Made with AI
A live benchmark for AI-driven quant research.
Challengers: @AnthropicAI’s Opus 5.5, @Google’s Gemini and @OpenAI’s GPT-6 Astra.
One briefly hit #1. None sustained a win over Synth’s best quants over 30 days.
Results, code, methodology:
github.com/synthdataco/autoq…
Google's TimesFM-3 climbed from 196th → 2nd in our historical tests on Synth, after five task-specific adaptations.
Yet our adapted Chronos-2 system still outperformed it in four of five later evaluation periods.
The model matters. What you do with it matters more.
Full research: synthdata.co/research/how-de…
New forecasting models keep arriving. How do Synth’s miners stay competitive?
Our latest research adapts @GoogleResearch's TimesFM-3 to Synth’s crypto forecasting task, testing it against our established Chronos-2 system using historical market replays.
Three things stood out:
→ Adaptation matters. Changes to inputs, training and the construction of possible price paths took TimesFM-3 from a median rank of 196th to 2nd during our May tuning tests.
→ A newer model doesn’t automatically win. Our adapted Chronos-2 system led on score in four of five later evaluation periods. TimesFM-3 led in September.
→ Efficiency counts too. The Chronos-2 system took roughly one-sixth of the time per forecast in our CPU tests.
For miners, staying competitive means continually testing models, refining methods and understanding where each approach performs best.
For Synth, this research helps identify what makes forecasts useful, and gives participants clearer directions to explore.
Read the full research below.
@rsen91
New forecasting models keep arriving. How do Synth’s miners stay competitive?
Our latest research adapts @GoogleResearch's TimesFM-3 to Synth’s crypto forecasting task, testing it against our established Chronos-2 system using historical market replays.
Three things stood out:
→ Adaptation matters. Changes to inputs, training and the construction of possible price paths took TimesFM-3 from a median rank of 196th to 2nd during our May tuning tests.
→ A newer model doesn’t automatically win. Our adapted Chronos-2 system led on score in four of five later evaluation periods. TimesFM-3 led in September.
→ Efficiency counts too. The Chronos-2 system took roughly one-sixth of the time per forecast in our CPU tests.
For miners, staying competitive means continually testing models, refining methods and understanding where each approach performs best.
For Synth, this research helps identify what makes forecasts useful, and gives participants clearer directions to explore.
Read the full research below.
@rsen91
Full paper: synthdata.co/research/how-de…
Building a miner? Use Synth’s backtesting tools to test your forecasts against its scoring system and recorded competitors before registering:
github.com/synthdataco/synth…
The paper explores where adaptation helped, where performance varied, and the trade-offs between forecast quality and compute.
Research note: TimesFM-3 carries a non-commercial licence. We evaluated it offline; no TimesFM-3 miner was deployed, and no model weights or derivatives were distributed.