Hi, So the issue is historic forecasts change when data last data input from a time series is deleted. So data is 1500 data point time series starting with 1.31238 ending with 1.32875 varying within 1.3 to 1.45. I prepare 5 lag bins which are converted to directional bins. So if the next data point is less than previous point the bin loads with a 1. If greater the bin loads a 2, if no change the bin loads a zero.
So I run in Jupiter Notebook the following script: (This follows the bin preparation scripts.)
In [108]: from sklearn.neural_network import MLPClassifier
In [109]: model = MLPClassifier(solver='lbfgs', alpha=1e-5,
hidden_layer_sizes=3 * [200],
random_state=5)
In [110]: %time model.fit(data[cols_bin], data['direction'])
The results are fine.
If I then delete the last data point and run the entire 100 lines of script, I get similar results but several of the 1499 predictions are different to the predictions from running the 1500 data points. varying from 6-8& differences or 120-140 differences.
The forecast is either 1 or -1.
It seems a bit fatal to then run the forecast every time a new data point is available if every time the historic or previously generated predictions vary to such an extent.
How do I continue calculating accurate forecasts ( generally 56-58% accuracy) without the historical predictions changing?