yevchyk/dancing_gorizon

Dancing Horizon is a machine learning system for generating trading signals on crypto and tokenized stocks like MU or PLTR. Instead of predicting a simple up or down at one fixed time, it answers a more specific question: if you opened a trade right now and closed it in some number of minutes, what is the probability it would end up profitable after fees. It answers that question for many possible time windows at once, producing a curve of probability over time rather than a single guess. Underneath, the system watches roughly 200 to 300 assets every 5 minutes, turns each asset's recent price and volume history into a set of around 300 numbers, and feeds those into two machine learning models built with a library called CatBoost, one predicting upward moves and one predicting downward moves. These models are described as symbol blind, meaning they look only at price and volume patterns rather than which specific asset it is, which is why models trained on crypto data are said to also work reasonably well on stocks they were never trained on. The project author is upfront that what matters is not how good the model looks in a backtest, but whether its stated probabilities actually match reality. So every prediction is checked against a calibration table showing the real historical win rate for that confidence level. The system only acts on the highest confidence predictions, described as the tail of the probability distribution, and does nothing on days where no such high confidence signal appears. The full pipeline includes fetching live market data from OKX and Binance, building features, training the models, running backtests, and optionally executing real trades, defaulting to a shadow mode that only logs signals without placing real orders. The author openly describes the code itself as messy, written quickly across many late night sessions, though the results reported are said to hold up regardless. The project is written in Python and released under the MIT license.