05
BTC News Reactions
A self-collected dataset joining multilingual crypto news to forward Bitcoin price reactions, from one minute out to twenty-four hours. Published on Hugging Face.
publishedHugging Facedataset · NLP · time series
// evidence
- observations
- 120,981
- horizon
- 1m – 24h
- hosted
- Hugging Face
- collection
- self-collected
// What is in it
120,981 observations. Each pairs a news item with what the Bitcoin price did afterwards, measured across horizons from one minute to twenty-four hours.
The collection is the point. Most dataset entries in a portfolio are someone else’s download.
// What this kind of dataset has to get right
- Availability, not timestamp
- What matters is when a story was actually readable, not the time printed on it. Training on a 09:00 article to predict an 08:55 move is look-ahead bias in a clean shirt.
- Time zones
- An Asian market, an exchange on UTC and a newsroom on local time is the quietest source of error in the entire pipeline.
- A time-based split
- A random split on a time series disqualifies the result before anyone gets as far as reading the metric.
// How it should be judged
By prediction quality — AUC, information coefficient, hit rate against a baseline that says the next move looks like the last one.
Not by backtest returns. A returns curve with no transaction costs in it reads as a mistake rather than a result.