A new financial dataset - FIN9.

In the stock price curve, there are often three actions involved: "buy", "sell" and "no action". These three actions are often carried out when the stock price trend curve has some specific characteristics. In this paper, we select 9 feature images (three "buy", three "sell" and three "no") make a dataset. The production process of the FIN9 dataset is shown in Figure Firstly, the feature curve should be selected. We selected nine types of feature curves, with three each for "buy," "sell," and "no". Then, the characteristic curve needs to be modeled, that is, fitted with a suitable piecewise linear function. Finally, we also need to convert the fitted image into a .ubyte file for network training and testing. In the training set, each type generated 6,000 28×28 black and white pixel images for a total of 54,000 training images. In the test set, each type generated 1,000 28×28 black and white pixel images for a total of 9,000 test images.