MinMaxScaler when LSTM predictions fall outside of training range?

Question

I am using MinMaxScaler on my training set and applying the transformations to my test set and inverse_transform to my model’s outputs. If this were, say, a stock prediction problem, my training set may have values between 1-200, but in the most recent time steps, the values are toward the high end of that range.

How can I handle future time-steps that should be predicted to land above 200 (since activation function outputs values between 0-1)? I guess my question is, is there a way to manually (or through some “rule of thumb”) specify a maximum value for the training set? So in this case, would it be possible to use MinMaxScaler to use 0-500 as the min-max values to transform to 0-1, that way values in my test set or predictions can be inverse_transformed accordingly? Is this an out-of-the-box scikit-learn feature? What’s the most efficient way to implement this, especially if this has to be done independently for multiple features?

Answer 1

Inverse-transforming with MinMaxScaler should be capable of producing something outside of the training data's range. It seems that, in your use case, using a final activation that lands in $[0,1]$ might not be appropriate. Even if you transform the training data to land in, say, $[0,0.7]$, applying a sigmoid or some-such on the final layer seems to lack motivation.

As to the question of whether it can be done: yes, just with a little roundabout thinking. You can't specify what you want the scaler to think your data's max and min are, but you can specify the output range you want (parameter feature_range), which amounts to the same thing.