Multiple Backtest Executions to Train Neural Network

It should technically be possible using a combination of ObjectStore and lean-cli to run backtests, but is it really necessary? It would be much simpler to simply pull some historical data and train the DNN in one run (perhaps iteratively in some loop).

Adam, the lean-cli is what I was thinking about using in the backtest runner I was referring to. It would be simpler if I could train on historical data alone but the problem requires that the weights are updated, e.g., via backpropagation, based on the results of the trading algorithm and not on historical data.  Historical data is only used as inputs to the algorithm.

If there are thoughts on how to accomplish the above without having to code outside of the framework, please let me know.

If the algorithm itself is simple enough, then one way that may work is to just simulate a quick and simple vectorized backtest inside each training loop. In psuedo-code, this will look something like

for n in 1,2,...,N:
    # 1. Simple vectorized simulation of your strategy
    # 2. Compute some metrics or whatever the objective function is here
    # 3. Update the weights of your DNN

Now if the algorithm has a bunch of moving parts like daily changing universes, complicated order logic, etc, then this will be pretty messy and you may have to automate some scripts using lean-cli.

I may have other ideas if you have more details (or a simplified example) of what it is you are trying to do

Adam,  Thank you for your input above.  I solved the issue by putting the neural network logic in the Initialze and OnEdnOfAlgorithm methods, saving the network and other needed variables in those methods, and using a bash script to run ‘lean backtest <PROJECTNAME> -d’ in a for loop.