S-TIER repo
Kronos — open-source foundation model trained on 12B candlesticks
100k ft — What is this?
Kronos — open-source foundation model trained on 12B candlesticks
50k ft — Why care?
Survey of zero-shot time-series foundation models (Google TimesFM 2.5, Amazon Chronos, Salesforce Moirai, IBM Tiny Time Mixers) that mostly fail at stock prediction, then Kronos: an open-source model pre-trained on about 12 billion K-lines from crypto and other markets that beats prior models on financial forecasting. Code on GitHub and Hugging Face from shiyu-coder.
10k ft — Key takeaways
• Generic time-series foundation models barely beat simple autoregressive baselines on stocks
• Kronos is domain-pretrained on candlesticks and is open source
• Pair any forecast with portfolio optimization, risk neutralization and cost control
• link: https://github.com/shiyu-coder/Kronos
5k ft — Tools & techniques
Hugging Face
algorithmic trading, fine-tuning, time-series foundation model, zero-shot forecasting
1k ft — Deep-dive / try it
There might be an AI model available to everyone that actually beats the market! Developed as a specialized foundation model for financial markets, “Kronos” was designed to speak the language of market data. Its training helps it to recognize complex market patterns across different assets and timeframes!
#newtech #aitrading #aifinance #newai #gigaqian
Transcript excerpt
gotta be honest I kind of hope this video doesn't go viral because a lot of you saw my video yesterday and asked me hey can I access that system and beat the market and it got me thinking and I might have found something I'm giga chin and as it turns out I might just be a man of the people I guess we'll find out I assume a few of you know that Google just quietly dropped a model it's called times FM named that because it's a time series foundation model as expected social media has been filled with people shouting it can predict anything but so can I and that doesn't mean I'm gonna be right and it certainly doesn't mean those predictions are going to have value I'm not saying it's a bad model like the old times FM the new one 2.5 has over 200 million parameters but this one has over 400 billion training points it's very good at some tasks but of course researchers wanted to test it on the biggest task arguably the hardest and most valuable predicting tomorrow's stock prices because there's a race going on this race is about something called zero shot forecasting which refers to predicting the future values of a never-before-seen time series without retraining the participants in the space really are a who's who of heavyweights Google has times FM Amazon has Kronos Salesforce has moirai IBM has tiny time mixers the list goes on obviously there was a lot of information here let me know if you want it but for now I'm gonna skip ahead to the point moirai is generally the best…
Appears in
More from @gigaqian
- B-TIER Xiaomi AI Cube — 80GB unified-memory box running 120B models locally
- B-TIER Moderna–Merck personalized mRNA cancer therapy — first positive Phase 3
- A-TIER Qwen 3.8 27B — local agentic model that fits on one 24GB GPU
- B-TIER Grok Bot — persistent AI workers sharing one cloud computer
- A-TIER Evo-designed bacteriophages — AI genomes that beat human gene swaps