Deep Learning in Characteristics-Sorted Factor Models

https://arxiv.org/abs/1805.01104

Example Data Description

1975 - 2017 monthly data (516 months)

• "realMKT.txt": a 516*4 matrix of Fama & French factors. The columns are Mkt.RF, SMB, HML and MOM.

• "realstock_return.txt": a 516*1000 matrix of stock returns, which contains monthly excess returns of 1000 stocks.

• "realZ_sample.txt": a 516*2000 matrix of stock charcteristics. The columns are size of stock 1, size of stock 2, ..., book to market ratio of stock 1, book to market ratio of stock 2, ...

• "realportfolio_return.txt": a 516*49 matrix of 49 Industry portfolio returns.

Usage of "dl_alpha" function

Input

• "data": a dictionary of all kinds of data. The keys are

  • "characteristics"
  • "stock_return"
  • "target_return" (i.e. portoflio return)
  • "factor" (i.e. benchmark factors)

• "layer_size": a list of hidden layer size where the last element is the number of deep factors to be constructed. For example, [32,16,8,4,2].

• "para": a dictionary of all training parameters. The keys are

  • "split": the way to split data into training set and test set. See "data_split" function.
  • "train_ratio": the proportion of training set if "split" is "future", between 0 and 1.
  • "batch_size": batch size for training. For example, 32.
  • "train_algo": optimization method. For example, tf.train.AdamOptimizer.
  • "learning_rate": parameter for "train_algo".
  • "activation": activation function when constructing deep characteristics. For example, tf.nn.tanh.
  • "epoch": the number of epochs for training.
  • "Lambda": tunning parameter for pricing error regularization
  • "Lambda2": tuning parameter for weight matrix regularization

Output

• "factor": deep factors

• "deep_char": deep characteristics