When creating Deep Learning Algo Trading systems based on history price data of a certain time period you can choose between several approaches to establish a model creation process:
- Splitting the complete history in exactly three sets (training, validation, test) and creating one model which you will use for your trading (Spoiler: Bad idea)
- Iterating over different timeframe sizes for the three datasets (training, validation, test) and this way only using a certain percentage of the available history to create exactly one model (as in 1.)
- Based on a fixed size of the three dataset timeframes (for training, validation, test) you create a shifting series of models and rate the series result instead of rating the result of a single model
- Based on a fixed size for the test dataset frame you create a series of models with variable training and validation dataset timeframe sizes. As size for the current training and test dataset timeframe size you select the training / validation timeframe combination with the highest validation-test result correlation in the previous period. The test dataset timeframe size defines the rhythm of model deployments.
The general idea behind this model production process evolution is the fact that markets change, trends change and players adjust their systems constantly. AI system developers have to implement this ability of constant adujstment in their lifecycle management. The less fixed hyperparameters you use the harder the model creation process is to handle, but the results improve significantly.