Earnings Prediction with Deep Leaning

In the financial sector, a reliable forecast the future financial performance of a company is of great importance for investors' investment decisions. In this paper we compare long-term short-term memory (LSTM) networks to temporal convolution network (TCNs) in the prediction of future earnings per share (EPS). The experimental analysis is based on quarterly financial reporting data and daily stock market returns. For a broad sample of US firms, we find that both LSTMs outperform the naive persistent model with up to 30.0% more accurate predictions, while TCNs achieve and an improvement of 30.8%. Both types of networks are at least as accurate as analysts and exceed them by up to 12.2% (LSTM) and 13.2% (TCN).

Motivation

Can we beat analysts with deep learning data based prediction?

Time Series Model

• Goal: predict earnings per share (EPS) of the next quarter for one firm
• Use information of the last $\beta$ quarters of the same firm

• Use skill score (SS) as a relative comparison value:
  ${\text{SS}}_{\text{MSE}}=1-\frac{\text{MSE}(m)}{\text{MSE}(\text{base})}$
• $\text{MSE}(m)$ is MSE of the own model $m$ (LSTM, TCN)
• $\text{MSE}(\text{base})$ is the MSE of the comparison model
  • Persistent model $pa$
  • Analyst forecast $a$
• Model under consideration is better (worse) than reference model if skill score is greater (less) than 0

Data Preprocessing

• Two datasets:
  • Accounting data from Computstat Quarterly
  • Daily stock market price and return data from CRSP Daily Shares
• Use about 20 important parameters
• Scale values using the totals assts $atq$:
  $x^{\prime }=\frac{x}{max\{1,atq\}}$
• normalize values:
  $z_i^{\prime }=\frac{z_i-\overline{z}}{\sqrt{\frac{1}{n}\sum _i(z_i-\overline{z}{)}^2}}$
• remove the outer percentiles of EPS, which are partially incorrect
• fill small data gaps with linear interpolation

Deep Neural Networks

• Use two types of neural network layers suitable for time series: LSTM and TCN
• Long-term short-term memory (LSTM)
  • Belongs to the recurrent neural networks
  • Consists of three gates (forget, input, output)
  

  Illustration of LSTM cell

• Temporal convolution network (TCN)
  • special kind of convolutional neural networks (CNNs)
  • one dimensional time series input
  • causal convolution (do not use future information)
  • dilation to cover a longer period of time
  

  Dilated causal convolution

Experimental Analysis

• Use two datasets:
  • A for choice of proper architecture and meta parameters
  • B for the final independent experiment with selected architecture
• Training: predict EPS values in the period 2012 to end of 2016
  • batch size = 1024
  • epochs = 1000
  • dropout rate = 0.3
  • window size = 20 (last 20 quarters and days)
  • optimizer = adam
• Selected two representative architectures

  LSTM architecture

  

  TCN architecture

• Testing: the following half year after the training
• Different experiments for financial (onlyfin) and non financial firms (nofin)
  • They behave significantly different
  • The following table shows different results on dataset A
  

  Table 1: Selected architectures and parameters for three groups of companies: financial (onlyfin), non-financial (nofin), and all.

  • results of best architectures on independent data set B:
  

  Table 2: Results on dataset B of optimal architectures and parameters grouped by financial sector affiliation.

Conclusion and Outlook

• LSTM and TCN models outperform the persistent model and analysts prediction regarding the MSE
• In the future we want to extend the experimental analysis
• Predictions could be used by broker firms as a supplement to the analysts predictions