It can always be a fascinating dream to predict the next day price of a stock by looking at the stock’s price curve. However, by thinking it through, I doubt the meaningfulness of do so, even with a very powerful technology, deep neural network.
A stock’s price, as time goes, rises or falls, from minute to minute. To view it abstractly, one can think of it can a continuous curve. A prediction task usually shifts a time window from past time toward current time. In each shift of the window, the oldest value is removed, and a newest value is added. In this process, one problem happens, that is the subsequence of any window is very similar to the subsequence of the next window. And next day’s price can either rise or fall, seemingly randomly. So, the input dataset is like assigning different labels to the same input data. Hence, the best that any model can do it to make a random guess on whether the next day’s price will rise or fall.
Some people have made some videos on predicting stock price using LSTM, or other models. As long as they use the stock price history to predict price of the same stock, it will not work, just because the inputs do not distinguish labels.
However, does it mean that stock prices cannot be predicted? Asserting this is still too early. There can be some relationships among various stocks, so there is still some possibility that one stock’s price can be predicted by other stocks.
Leave a Reply