Hands-On Machine Learning for Algorithmic Trading

Model training and evaluation

For illustration, we will create a document-term matrix containing terms appearing in between 0.5% and 50% of documents for around 1,560 features. Training a 15-topic model using 25 passes over the corpus takes a bit over two minutes on a four-core i7.

The top 10 words per topic identify several distinct themes that range from obvious financial information to clinical trials (topic 4) and supply chain issues (12):

Using pyLDAvis' relevance metric with a 0.6 weighting of unconditional frequency relative to lift, topic definitions become more intuitive, as illustrated for topic 14 about sales performance:

Sales ...