Field reports from building Swing Finder. What worked, what didn't, and what the data actually says about trend-line analysis once you measure it.
A model that scores 62% on a held-out test set is fine. A model that scores 62% on a test set you're sure was actually held out, with metrics you didn't pick to flatter the result, is much better. The methodology, the bugs we hit getting there, and the full numerical receipts for all twelve models on 2.58 million historical signals each.
Read it →If "will this line hold?" is the wrong question, what's the right one? We trained twelve neural networks on 2.58 million held-out historical signals. Sixteen experiments, three reboots, and one architecture that finally worked. Headline finding: the model is right 70 to 81% of the time when it's confident, and correctly hovers at chance in the middle deciles. It knows what it knows.
Read it →You hear it everywhere — technical analysis is horoscopes for the financially curious. I built a 1.2-million-event statistical lattice to settle the argument. The skeptics are not entirely wrong. They are also not entirely right, and the part where they aren't right is small, weird, and worth knowing about before you put any money on a chart pattern.
Read it →