
I have been doing time series forecasting for a while. I have read too much information and have used almost all existing popular models by myself. The conclusion is that if the data is potentially regular, it can be predicted, otherwise everything is nonsense.
Look at the many answers in the answer. Many of them are the same as when I first came into contact. I guess I can use any lstm or something. The effect is surprisingly good. Yes. I have already thought about it this way. This is too simple. I can buy stocks. Let me give you an example.
时间序列预测股票的困境
Borrowing this picture, I can’t predict that it’s accurate enough. Maybe everyone hasn’t figured out one point. This picture is drawn based on the test set. For example, three points predict one pointthen[x1,x2,x3]Actual data prediction[x4]and then[x2,x3,x4]prediction[x5]Note that the difference here is no prediction x5* uses the real value x4 instead of the previous round of prediction x4* so that the predicted value in each round is the real value instead of The prediction value of the previous round so that all the prediction points are based on the actual real value without adding new prediction values. What deviations can the picture drawn like this be? In actual situations, we predict that the subsequent rolling prediction is[x2,x3,x4*]prediction[x5*],[x3,x4*,x5*]Predict x6*, so it keeps scrolling because every time a new forecast value is brought in, the error will get bigger and bigger in the later stage. If you don’t believe it, you roll from 400 to forecast 500. I dare not say this picture Any model can be predicted as shown in the figure above.
So don’t be fooled by the test set. The test set predicts the next point based on the true value, so the error will not be too big. But in reality, where the true value can only be continuously rolled to have subsequent predictions
So this picture wants to predict later, there is no law at all, and the fraudulent model is also misunderstood. Yes, there is no solution. There is no way to predict and have no solution.
Let’s go back to the time series forecasting. At present, there is no solution for the time series ofwithout any regularity and no way to predict.
But you said to use lstm to predict the sin function. Isn’t it easy to catch it? So the obvious law is that lstm can’t predict what you eat? Don’t lstm, arima, etc. will have the same effect
For this kind of irregular timing, the way I can do it now is to smooth it and then decompose it.
时间序列预测股票的困境
Like this, it is forcibly decomposed into trend cycles plus noise, and then the trend cycles are predicted separately, and then combined together. It will be better to solve this problem. Facebook’s prophet is this idea. You can use it directly.
Finally, for irregular timing, don’t use any algorithm without doing any processing. It’s all nonsense and a waste of time. Don’t be fooled by the appearance of the test set.
And for periodic and regular timing, any algorithm will achieve good results. That’s right. 
