This AI Learned from 12 Billion Candles. Can It Trade Gold?

A CodeTrading video testing Kronos, an academic transformer model trained on 12 billion candles, to forecast gold prices across five timeframes over 18 months of out-of-sample data.

Key Takeaways

Kronos treats candlestick charts as a language, predicting future candles much as ChatGPT predicts the next words.
A single forecast proves nothing; statistical validation over many forecasts and timeframes is required to judge any edge.
The test uses a strict out-of-sample holdout of 18 months to avoid tuning on test data.
The model is free and accessible on consumer hardware, making it easy for retail traders to experiment.

What the video covers

The video introduces Kronos, a ChatGPT-style foundation model trained only on candlestick (OHLCV) data, released by researchers from Singu University.
Kronos was trained on about 12 billion candles from 45 exchanges and is free, with code on GitHub and models on Hugging Face.
Available model sizes include Kronos mini, small, and base; Kronos large is not public.
The smaller free models run on a normal gaming graphics card, and the research paper is available on arXiv.
Kronos has two parts: a tokenizer that converts each candle into a short discrete code, and a transformer that predicts the next candle codes autoregressively.
Input is typically the last N candles (here 90) as an OHLCV table; output is a forecast table for the next 10-20 candles.
The experiment uses six years (2020-2026) of one-minute gold data, aggregated into 5-minute, 30-minute, 1-hour, 4-hour, and daily charts.
Kronos-mini (about 100 million parameters) is used out of the box without retraining, and the last 18 months are held out as a test set.
The presenter stresses that single forecasts are meaningless and thousands of forecasts are needed to assess statistical edge.
The headline claim of almost 40% gold trading returns in 18 months is presented as the motivating hook for the test.

Chapters

What is Kronos and who created it?

Kronos is a transformer-based model trained only on candlestick data, released by researchers from Singu University as an academic project. It is free, with code on GitHub and models on Hugging Face.

How does Kronos turn price charts into predictions?

A tokenizer converts each candle into a short discrete code, and a transformer reads the sequence of codes and predicts the next candle. The predicted candle is appended and the process repeats to draw a multi-candle forecast.

Why does the presenter say one forecast is not enough to judge the model?

A single forecast can easily be wrong, as shown when Kronos predicted a fall that did not happen. Thousands of forecasts across timeframes and an out-of-sample period are needed to see whether the model has a statistical edge.

Full Transcript — Download SRT & Markdown

Speaker A

This AI made almost 40% trading gold in 18 months. Imagine a chat GPT model built only for candles.

Speaker A

It reads a price chart and predicts the most likely next candles.

Speaker A

And that's Kronos.

Speaker A

It was released by researchers from Singu University.

Speaker A

So this is an academic research project.

Speaker A

They trained it on 12 billion candles from 45 exchanges, and it's free.

Speaker A

The code is available on GitHub, and the models are on Hugging Face.

Speaker A

And all the details related to the architecture of the model and how they pre-trained it, on which data, etc., is also available on the UH GitHub repository with the available models.

Speaker A

So we can see that we have the Kronos mini, the Kronos small, and the Kronos base, while the Kronos large is not published for us, for the public users.

Speaker A

But that's okay.

Speaker A

The free versions are big enough for our test, and they run on a normal gaming graphics card.

Speaker A

The authors also published their study on Arxiv.

Speaker A

So if you are interested in the research part, you can download the full paper as a PDF file.

Speaker A

So today I'm putting Kronos to the test on gold.

Speaker A

I ran almost 1,000, or 100,000 forecasts, with five time frames and six years of gold data to see what it can really do.

Speaker A

ChatGPT learned the language of text.

Speaker A

Kronos learned the language of candles.

Speaker A

Each candle is like a word, or it's treated like a word.

Speaker A

A chart is like a sentence.

Speaker A

And just like ChatGPT tries to write the next words in a sequence of words, or in a sentence, Kronos writes, or estimates, the most probable next candles.

Speaker A

Kronos has two parts.

Speaker A

The first one is a translator.

Speaker A

It turns each candle into a short code, like a word.

Speaker A

First the rough shape of the candle, then the details.

Speaker A

Part two is the brain.

Speaker A

It's a transformer, the same kind of network, or neural network architecture, as ChatGPT.

Speaker A

It reads the chain of candle codes and guesses the next one.

Speaker A

Then it adds that candle to the stack of available candles and tries to guess again for the future candle.

Speaker A

So candle by candle, it draws the future, or it tries to estimate at least the future.

Speaker A

So what goes in and what comes out?

Speaker A

What do you give the model?

Speaker A

We usually give it a table of past candles with open, high, low, close, and the volume columns.

Speaker A

One row per candle.

Speaker A

For example, in this experiment today, I gave it the last 90 candles, or the last 90 rows, every time.

Speaker A

I needed an estimation for the next future 10 candles.

Speaker A

You also tell it the times of the future candles you want to do the estimation for.

Speaker A

And what comes out?

Speaker A

The same kind of table for the future candles.

Speaker A

You get the open, high, low, close, and the volume for the next UH 10 candles, or 15, or 20 candles, for example, depending on your parameterization.

Speaker A

Now, here's the UH real one.

Speaker A

Kronos said, for example, the gold would fall in one of the estimations I tried, and it actually went up.

Speaker A

One forecast means nothing.

Speaker A

This is why we will need thousands of forecasts to check if statistically this model provides an edge in our trading systems.

Speaker A

Now, for gold, I used six years of one-minute gold prices from 2020 to 2026.

Speaker A

That's around 2.1 million rows.

Speaker A

From them, I built five charts: 5 minutes, 30 minutes, 1 hour, 4 hours, and the daily.

Speaker A

So, different time frames.

Speaker A

I used Kronos-mini, which is about 100 million parameters, with no retraining.

Speaker A

I used it as is, out of the box.

Speaker A

And I locked away the last 18 months as a test period.

Speaker A

So we don't touch this data.

Speaker A

I didn't tune anything on it.

Speaker A

This is our out-of-sample data, just to be able to validate the model.

Speaker A

The trading rule is simple.

Speaker A

I used the model in a trading system, somehow, to be able to backtest it.

Topics:Kronosgold tradingcandlestick AItransformertime series forecastingalgorithmic tradingbacktestingHugging Facefinancial foundation modelXAUUSD

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

App Store Google Play

Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools