It seems like predicted price is off 1 day on IBM or maybe i missunderstood something about time-series-forecasting-pytorch

[deriklogov]

to be clear, to me it looks like next day prediction price is very close to previous day actual price

[cdiener1]

Same problem. I don't get it.

[ujlaki15]

I am just leaving a comment here to catch your solution. I am also working on this problem right now.

[pdellis111]

how did you go with this? I am looking at the same problem.

[ujlaki15]

Ok, my idea is that the machine is cheating. It gives you back a value very close to the previous day, since it should be the closest one to the real value without doing prediction.
I used different networks and the result was usually the same. Doublechecked every variable, label, etc. That wasnt the problem.

[egorpes]

Perhaps a little late, but...
Making price predictions based on past prices is a poorly feasible task (I think generally impossible). The first thing that stands out is that the predictions are perfect (but "shifted"). The data link is not good enough for the model to make such good predictions.
If you want at least some predictions based on previous courses, you can try the step not for a day, but for a week or a month. In such a case, a stable market can have a normal result (when was the market last stable?).
If you still want to analyse (whether for a day or for a week does not matter), then add at least a hundred economic parameters to the model. The problem with learning is not writing code (with libraries like PyTorch or TensorFlow it is not very difficult), but preparing the data. Alphavintage offers a handy API, but you need data, for example, from any statistical bureau. Be sure to watch all reports and the US economy (there are economic calendars, although APIs are paid for historical data almost everywhere), a little for China, Japan and the EU.
In general, machine learning is not a panacea. If you have several trillion parameters, then maybe it will. Right now you can add a hundred parameters and see if there is a prediction (it does not matter if it is inaccurate). I have not tried it, but you can use fingpt, for example, to analyze non-numeric data. Then give it your predicted numbers and listen to the advice for the next day.
UPD. Forgot to say. Do not analyze absolute values (even normalized ones). That is completely useless, in my opinion. Analyze the change in magnitude. For example: value_rel = (value_cur - value_prev)/value_prev

[nathanfranke]

@egorpes You can also simplify (current - previous) / previous to current / previous - 1.0