Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Generating Probability Distributions for Future Stock Prices

Our goal is to build a deep learning model that reads historical data and outputs detailed probability distributions of future stock prices, rather than simply the predicted price. The input to our algorithm is of the shape (n, f, d), where n is the number of stocks, f is the number of features per stock per day, and d is the number of days of history. The inputs progress through convolutional layers followed by fully-connected layers and a softmax output. The resulting model predicts the future percent change in price on the next day for each stock, and assign probabilities to each range of percentage change (the boundaries between each range being -4%, -3%,-2%,- 1%,0%,+1%,+2%,+3%, and +4%).

Dataset & features

The daily stock price dataset from the Hong Kong exchange was gathered from Quandl. This time-series data is discretized per day, and preprocessing included cleaning the data to fill in missing values with the previous day’s value. Below is an example of the values for one stock indicator:

Date Price Bid Ask High Low Previous Close Share Volume Turnover
2020-01-02 1.840 1.830 1.850 1.850 1.830 1.830 2246.0 4155.0
2020-01-03 1.830 1.830 1.850 1.840 1.810 1.840 26.0 48.0
2020-01-04 1.900 1.900 1.910 1.920 1.820 1.830 5395.0 10350.0


The initial method explored is to take the time-series data and apply various deep learning techniques to predict the future price. The target value is the historical price of a stock on a certain day, and the features used to predict the target value are the daily features from the time-series history from the preceding year. In order to normalize this target value, it is represented as a percent change from the day before, rather than the nominal price. The loss is the mean-squared error between the actual (target) and predicted value. After this method, we then introduce the notion of predicting the price probability distribution, so the target y value changes from a being price change, to being a 10 dimensional one-hot vector to indicate which range the price change is contained in.

Experiments & results

Experimentation began with simple fully-connected layers operating on price history only, to determine a baseline. These techniques used a learning rate of 0.0001, and ran for 100 epochs. The experiments were run with a 1-year history on a small set of 200 stocks, to speed up model structure and hyperparameter experimentation. Early results showed that a model with 3 fully-connected layers would outperform models with fewer layers, and that the tanh activation function between layers was suitable. Next, experimentation began on convolutional layers. At first, these convolutional layers operated on price history only. A learning rate of 0.0005 was used to speed up training so that comparable results could be achieved in 100 epochs. Various kernel sizes, channel counts, convolution layer counts, and max-pool sizes were used, and 2 convolution layers with max pool of small kernel size was found to be suitable. To further enhance the model, seven additional time-series features were introduced: bid, ask, P/E ration, daily high, daily low, volume, and turnover. Along with price, we had 8 features, and we fed the expanded feature set into a convolutional model from previous experiments. This new model outperformed our previous models:


Below are the training, development, and test set losses for the approach with convolutional layers with max pool, followed by fully connected layers. The train/dev/test split is 60%/20%/20% in stocks.

Training set loss Development set loss Test set loss
0.0184 0.0262 0.0363

Finally, we added a softmax layer to the output, and reformatted the target value to predict the range in which the price change would be contained. By analyzing our baseline models from above, we chose the structure of this model to be the best-performing model that we had found before, and we attached a softmax layer to the end of it. This final model has a first convolutional layer of 20 channels, kernel size of 7, stride of 1, with a max-pool of size 3. The second convolutional layer had 1 channel, with kernel size of 7, stride of 1, and a max-pool of size 3. The third layer is a fully-connected layer of size 84, and the fourth layer is fully-connected of size 10, resulting in a softmax of size 10. This softmax-based model ran for 500 epochs, with a training rate of 0.005 using the Adam optimizer.

Below are 5 example stocks from the test set, where the price history is shown on the left as a reference, and the model’s predicted price distribution is shown on the right. Additionally, the training, dev, and test set losses are included:


Training set loss Development set loss Test set loss
0.0686 0.0826 0.0815

Conclusion & future work

We find that a deep learning model for predicting stock price can serve as a reasonable baseline model. Furthermore, using softmax to predict the discretized range of price change produces reasonable and richer price probability distributions. Our specific results showed a test set MSE loss of 0.0815 when predicting the correct magnitude of percentage changes. The validity of generating probability distributions in this way has several promising implications, such as being able to model more detailed aspects of risk and price movement. Other aspects of future work include expanding the number of features used, and extending this predicted probability distribution to generate simulated future price history.


  1. Feng, G., He, J., Polson, N. G. (2018) Deep learning for predicting asset returns. arXiv preprint arXiv:1804.09314.
  2. Chen, L., Pelger, M., Zhu, J. (2019) Deep Learning in Asset Pricing. Available at SSRN:
  3. Sirignano, J., Cont, R., (2019) Universal features of price formation in finan- cial markets: perspectives from deep learning, Quantitative Finance, 19:9, 1449-1459.
  4. Yu, P., Yan, X. (2020) Stock price prediction based on deep neural networks. Neural Comput Applic 32, 1609–1628.
  5. Qiu, J., Wang, B., Zhou, C. (2020) Forecasting stock prices with long-short term memory neural networks based on attention mechanism. PloS one, 15(1), e0227222.
  6. Naik, N., Mohan, B.R. (2019) Intraday Stock Prediction Based on Deep Neural Network. Natl. Acad. Sci. Lett.
  7. Yun, H., Lee, M., Kang, Y. S., Seok, J. (2020) Portfolio management via two-stage deep learning with a joint cost. Expert Systems with Applications, ISSN: 0957-4174, Vol: 143, Page: 113041.
  8. Quandl Hong Kong Exchange Daily Stock Prices Kong-Exchange


Generating Probability Distributions for Future Stock Prices






No releases published


No packages published