Why Spearman’s IC? What about accuracy?

Why RMSE Can Be Misleading for Crypto 📉

RMSE (Root Mean Squared Error) measures how far off a prediction's magnitude is from the actual result. It heavily penalizes large errors because it squares the difference between the predicted and actual values.

In most forecasting tasks, this is a good thing. But in crypto, a massive price swing (an "outlier") is often a real market event, not just noise. A model that correctly predicts a huge upward spike but is off on the exact price would be severely punished by RMSE, even though it provided an incredibly valuable signal.

Analogy: Imagine RMSE is like trying to predict the exact time a horse will finish a race. Being off by a few seconds is a big error, even if you correctly picked the winning horse.

Why Spearman's IC Is a Better Fit 🎯

Spearman's IC doesn't care about the exact price. Instead, it measures how well the model ranks the performance of different assets. It answers the question: "Did the model correctly identify which cryptocurrencies would perform best and which would perform worst?"

This is much more robust to outliers. An extreme price movement is simply treated as the #1 ranked event, and the metric checks if the model also gave it a top rank. This aligns better with trading strategies where getting the direction and relative strength right is more important than predicting the exact price.

Analogy: Spearman's IC is like simply picking which horse will win the race. You don't need the finishing time, you just need to get the order right to win your bet. 🐎

Regarding "90% Accuracy"

For a regression task like price prediction, "accuracy" isn't measured in the same way as a classification task (e.g., "spam" vs. "not spam"). Instead of a single accuracy score, we look for a model with a consistently positive and high Information Coefficient. A high IC proves the model has a genuine predictive edge in ranking future market movements, which is the core goal for this type of time-series data.