• Every bitcoin transaction
• Every second
• Full blockchain history
• Live mempool entries
• Huge TXs identified
• Round USDs identified

Yeah, it works on the logic that most transactions are in round dollar amounts, like $50, $100, etc, so there's an extra density of transaction outputs at those round amounts. By looking at where these peaks in the transaction outputs distribution are, you can estimate what the price would have been to make those sat amounts equal to round dollar amounts. I think it's a really cool idea, especially because you don't need to pull the price from any external data provider.

This doesn't try to predict the future at all. It just allows you to pretty accurately determine a given day's USD price (assuming the day is within their specified range). I don't fully grok it, but most of the magic is in a historical "stencil" in the shape of a bell curve that matches output patterns to price points. It then lines up this stencil with a bell curve of the output patterns on a given day and voila. What's cool about this is that if you need historical price data you can use this to deduce the price rather than rely on some external source. I guess it could break if onchain behavior for a day is anomalous and not accounted for by the model?