Are analysts’ Forecasts Reliable? A Machine Learning-Based Analysis of the Target Price Accuracy

Abstract

This paper examines the accuracy of target price forecasts made by sell-side analysts, focusing on predicting target price accuracy using machine learning approaches. Utilizing a dataset of target price forecasts for U.S. listed companies from 1999 to 2021, we employ ensemble methods and incorporate market-level, firm-level, and analyst-level information to predict target price accuracy in terms of target price errors and target price achievement. The long-short portfolio constructed based on our predictions significantly outperform the benchmark in terms of cumulative return and Sharpe ratio.

Keywords:

• target price
• analyst report
• machine learning
• ensemble methods
• portfolio

JEL CLASSIFICATION:

• G14
• G17
• G24

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 “Price targets are hazardous to investors’ wealth,” https://www.nytimes.com/2001/08/05/business/market-watch-price-targets-are-hazardous-to-investors-wealth.html.

2 According to Hao and Skinner (2022), most analyst reports state that the forecasted 12-month dividends have been discounted in the 12-month target prices. The target prices do not include dividends to be paid within a year. Moreover, the dividend payout ratio has been included in the firm characteristics.

3 The I/B/E/S database provides target price data dating back to July 1999.

4 Sensitivity: the “true positive rate,” or the percentage of positive cases the model can detect; specificity: the “true negative rate,” or the percentage of negative cases the model can detect.

5 The number of estimators “n_estimators” is reduced from 500 to 150 to mitigate the computation time.

6 For example, if there are 2024 target prices in a certain month, we include 202 stocks in both the long and short segments.

7 For gradient boosting regression model, we use squared error as the loss function to be optimized.

8 For gradient boosting classification model, we use binomial deviance (negative log-likelihood) as the loss function to be optimized.