I’m working on building a model to forecast demand for 1000s of different things. Price will be a variable included in the model, so after I’m done tuning my best model, I’d like to find “optimal” prices to use.

I created a couple of examples in Keras to test the idea and it seems to be working on a small scale. You can view my notebook here.

My method is somewhat simple

- Build a model with price as a variable
- Freeze the model weights
- Create a single trainable weight (weights for multiple items) and insert it right after the price input
- Also multiply that weight with the final output of first trained model that estimated sales to get revenue
- Train model (that single weight) to minimize the negative output (which maximizes expected revenue)

I was hoping anyone here could give some criticism or advice. My model will be a bit larger scale with way more inputs and I will probably be using a recurrent net as well since I’m forecasting time series. I’d like to avoid any pitfalls or statistical blunders that might pop up from naive model assumptions.

Thanks!