A ranking loss and a profit threshold

A paper posted this week asks a deliberately narrow question: will Bitcoin rise by more than 5% within the next seven days. It answers it with a multi-scale temporal convolutional network trained on on-chain, market and sentiment data spanning February 2018 to December 2025.

The architecture combines InceptionTCN blocks, channel attention, adaptive average pooling and a pairwise ranking loss, with dilated convolutions capturing horizons from one to four days. Against five baselines — an improved TCN with a gated recurrent unit, an LSTM, a plain TCN, XGBoost and a random forest — it reports an area under the curve of 0.6316 and a profit measure of 1.703, ahead of all of them.

The methodological choices are the part worth carrying into practice. The dataset is class-imbalanced, so the authors report area under the curve rather than accuracy, on the grounds that accuracy on an imbalanced binary problem mostly measures the base rate. And rather than take the conventional decision threshold, they select one optimised for profit, aligning model selection with the objective the model exists to serve.

An area under the curve of 0.6316 is modest in absolute terms and the authors do not dress it up. For a directional call on a seven-day horizon, modest and honestly measured is the normal shape of a real result; the papers that report much more usually turn out to have leaked something from the future into the training window.