Crossbreeding models means picking a number of models and then generating a few new ones, which have all the parameters being similar to the parent models, but which have their inputs mixed in a few ways.
When you cross-breed models, the system does the following:
- Pick a few “best” inputs, a few “worst” inputs, and a few random inputs from the overall list of inputs of all the parent models
- Mix them in a semi-randomized fashion. Also apply a few common feature engineering techniques to some of them.
- Create a few new models from that.
Provided that crossbreeding uses “feature importance data”, applying this technique requires all the parent models having the same model type, signal parameters and training data sets. That’s because this is the only case when comparing “feature importance” makes sense.
Crossbreeding models is useful in a few cases:
- You are feeling adventurous and want to experiment fast.
- You are following the workflow we have suggested.
- You don’t know what you are doing. In which case crossbreeding is akin to putting a few parts in a box and then shaking it aggressively, hoping for the parts to assemble themselves properly. That might work!
Crossbreeding is a really good way for experimentint fast. We have seen a few generations of crossbreeding producing models which were substantially better than the founding models.