Situation
I kept seeing machine learning trading projects online that showed a beautiful equity curve and no discussion of how it was tested. I was suspicious of that and wanted to find out for myself whether ML actually adds anything in markets or whether it mostly just fits noise.
Task
Build a full strategy end to end, but hold it to a standard where I would believe the result. That meant the validation mattered more to me than the models did.
Action
I built the strategy out of three signals that do different jobs.
Unsupervised clustering groups assets by how they actually behave rather than by their sector label, so diversification is based on observed correlation instead of what an exchange calls something.
Sentiment analysis over news and social data gives a directional prior.
Volatility forecasting sizes the positions, so exposure comes down when the model’s own uncertainty goes up.
The part that took the longest was the testing. I used walk forward validation instead of a single train test split, so the model only ever sees data that would have existed at the time. Then I put realistic transaction costs in.
Result
The pipeline works and the notebooks lay out each stage so the reasoning is visible.
The honest finding is that transaction costs killed most of what looked profitable. Several versions that produced a lovely backtest curve went flat or negative once I charged realistic costs per trade. The signals that survived were the slower ones that traded less.
That is the main thing I took away. The modelling was the easy part. The hard part was building a test I could not fool, and then accepting the answer when it was not the one I was hoping for.