Conversas em AI-Depot

• Stock Market data: I think data scaling incorporates future data...
I'm really sorry if that's a dumb question - I'm not a math expert. But I'm facing this problem:

Simple scaling of data to fit inside -1 to 1 values incorporates future data information, because values will be different if "future" values are different, invalidating stock market prediction.

Is this true? How can this be solved?

amcmr - 2 posts. Thursday 25 March, 13:36

• Scaling Nonstationary Data
You're right that such scaling would incorporate "future" information, but the bigger issue is that the data is nonstationary, meaning in this case that it drifts over time. Taking first differences or first ratios or some type of basic time-series analysis would likely improve your situation.

-Predictor - http://will.dwinnell.com. 100 posts. Thursday 25 March, 14:19

• Converting to percent change and decimal scaling
...Hmmm, I see. For stock market prices, I'll check if a better solution is to normalise data by replacing absolute price values with accumulated day ratio changes since day 1 of analysis, and then doing a fixed decimal scaling (http://www.netnam.vn/unescocourse/knowlegde/2-2.htm) to make values less than |1| in order to feed a neural network pool.

Depending on each particular system, maybe keeping always the same time frames of analysis is important.

Thank you, Predictor. btw, nice website.

Sem comentários:

Enviar um comentário