Principles of time series analysis by neural network models

External question — Data Science Stack Exchange Author: feynman Original post: https://datascience.stackexchange.com/questions/46765 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. I can understand that, in the case of speech signals, words are correlated, and therefore one should have a reason to believe that RNNs or LSTMs could predict future observations by running some complex algorithm with weights and an activation function. But for random digital signals like stock prices, why can the prices be predicted in any way? Tomorrow's price has nothing to do with the prices that showed up in the past.

External answer — Data Science Stack Exchange Author: Georg Unterholzner Original post: https://datascience.stackexchange.com/a/46788 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. According to the Efficient-market-hypothesis , you are right: the price movement of today does not affect the price movements of yesterday. However, this hypothesis describes an abstract assumption, which can be very useful for macroeconomic modeling but should not be considered as a comprehensive description of reality ( The same applies to other economic assumptions, like that of the Homo economicus for example. ). In fact, there is a lot of criticism regarding the efficient market hypothesis. The Cryptocurrency-Bubbles and the Dot-com bubble are most probably the best examples that price movements in financial markets are not pure random walks in many cases, but interfere with behavioral psychological effects. From my experience, I can assure you that many indicators are quite consistent: if you're trading in a growing market, the probability of an increase in prices is slightly higher than that of a decline ( even though there will be a tipping point sooner or later ). In addition, there are other indicators, such as volatility, which are relatively stable. Of course, you could try to manually determine all these indicators and include them as features in an ordinary DNN, but in practice, using an LSTM usually turns out to be easier and more accurate.