Earning surprise betting strategy performance
Earnings expectations and earnings forecasts are a central focus in the financial industry. Stock price adjusts according to a company’s ability to increase earnings as well as to meet or beat analysts’ consensus estimates. Betting on earning reports is a popular strategy which focuses on the earning surprise: the percent difference between the actual and estimated earnings per share. In this paper, we first backtest the historical performance of the earning surprise betting strategy and deem it profitable. Based on this conclusion, we investigate several machine learning model architectures to predict the earnings surprise using historical fundamental data along with Bloomberg Estimate EPS. We also backtest the strategy using the earnings surprise projected by different models and find that the attention-based model yields the best performance.
Bloomberg estimates data
Bloomberg’s estimates dataset delivers three types of key forward-looking measures: consensus estimates, analyst recommendations, and company guidance. Consensus estimates fields include all key Income Statement, Balance Sheet and Cash Flow measures from Sales, EPS to Net Debt and Free Cash Flow. Both Generally Accepted Accounting Principles (GAAP) and Adjusted fields are available for Net Income and EPS. For selected fields, measures of high, low, medium, four-week change, and number of contributions are also provided for additional context on the data distribution. For the purpose of this paper, we focus on the consensus estimate of quarterly EPS.
Bloomberg fundamentals
Bloomberg covers the entire financial reporting process of companies – from earnings to preliminary releases to full fundamentals. With fundamentals data on more than 85,000 companies (both active and inactive) starting from the late 1980s, this dataset is the backbone of Bloomberg’s Equity solutions. Bloomberg’s fundamental coverage includes current and normalized historical data for the balance sheet, income statement, cash flows statement and financial ratios, as well as industry-specific data for communications, consumer, energy, health care and many more.
Advanced machine learning techniques
We build a trading strategy based on the model predicted earning surprise. For each quarter from 2019 Q1 to 2021 Q4, stocks with top sixth earning surprise are included in the portfolio and held for one day. We use different architectures for earning surprise prediction, including linear regression, multi-layer perceptron, long short-term memory, and attention-based model. Among all these techniques, attention-based model yields best results.
Backtesting results
We test the performance of the benchmark strategy (a strategy purely based on historical earning surprise) and attention-model-based strategies from 2019 Q1 to 2021 Q4. Both benchmark and attention-model-based strategies show attractive Sharpe ratio and cumulative return. The attention-model-based strategy outperforms the benchmark strategy in terms of both cumulative return and Sharp ratio in the testing period.
Download the whitepaper to learn more about trading strategies using earning surprise.