Quantitative earnings forecasting has evolved over the last 25 years of my career in the quantitative finance space. I’ve been fortunate to participate in some of that evolution, including as head of research on the founding team at StarMine in the late 1990’s and early 2000’s where I designed the SmartEstimate product.
I’m sharing information from my published white paper, “TrueBeats for EPS and Revenues,” with 3 methods you can use to predict earnings and revenue surprises even more accurately.
The Old Methodology
The creation of the Institutional Brokerage Estimates System (I/B/E/S) in the 1970s—which combined forecasts from multiple brokers to create an equally weighted “consensus” – was the first game changer. This simple consensus remains common in many financial applications today, despite extensive academic and practitioner research demonstrating various aspects of consensus forecast errors in the ensuing decades.
Starting in the 1990s, academic and practitioner research started to show consensus forecast errors could be mitigated by considering who was providing the estimates, along with their track record, tenure, employer, timeliness, and other features. New and innovative firms such as DAIS Group, Zacks, and perhaps most notably StarMine, built analytics to come up with a better composite forecast of earnings – and earnings surprises – by using these types of analyst-specific features.
But few of these methodologies have evolved or improved in the last 15 – 20 years. They are limited by the range and quality of forecasts from the underlying analysts. However, there are cases when all analysts are similar and / or they do not provide the most significant source of forecast accuracy.
A simple weighted average of an analyst’s estimates based on individual characteristics will fail to take into account a wide variety of information. For example, from corporate earnings, including time series and cross-sectional effects.
The amount of information has increased, along with the data-science tools at our disposal, affording new ways to consider earnings-surprise predictions that may go beyond what is available in commonly used commercial models.
Below are several enhancements and alternate approaches we have tried in our research. Although there is now a wide variety of data available for corporate earnings, including the unstructured text from earnings call transcripts.
Here are 3 Effective Ways to Improve Earnings Forecasts:
1: Overweight Estimates Based on Analysts’ Ability to Predict Specific Items and the Recency of Their Estimates
Older “intelligent estimate” models such as StarMine’s SmartEstimate weight forecasts based on the historical accuracy of each analyst and the recency of each estimate. The most recent estimates from the best analysts receive the highest weight.
As mentioned above, differentially weighting forecasts based on the contributing analyst’s track record can result in a more accurate estimate of earnings or other KPIs. But one can improve the accuracy rate in several critical ways.
First, one can look at the historical accuracy of each analyst for each item individually. For example, some analysts may be better able to predict revenues compared to EPS, perhaps because of the varying quality of their estimates of expenses. A composite sales estimate, which is weighted by the analyst’s track record in EPS estimation, is more accurate than the consensus 51.8% of the time, whereas using the analyst’s track record in sales estimation results in a 56.6% accuracy rate. The same phenomenon is true for forecasts of cash flow, dividends, and same-store sales.
Additionally, a dynamic model can help determine weights for prior periods when measuring historical analyst ability. Although analyst ability is somewhat persistent over time, to the extent that it changes it may do so at a varying rate. Analysts who were accurate in one market regime may become relatively inaccurate after a regime change.
There are also observable seasonal effects in the data, where an analyst’s accuracy four, eight, or 12 quarters ago is especially important in predicting the individual’s accuracy in the current quarter. This is because some analysts are particularly good at prediction in certain fiscal quarters relative to others. In fact, the four-quarter-ago accuracy is nearly twice as important as the three-quarter-ago accuracy.
2: Follow the Trend in Beats and Misses for Each Stock and Its Peers
Earnings surprises do not occur in isolation; each company’s time series of prior earnings and revenue surprises contains valuable information on how efficiently the firm has beat expectations in the past.
The experience of a firm’s peers is also important, as earnings surprises in a particular industry tend to cluster in time. The simple way to define peer stocks is by using industry classifications. An improved method defines peers with overlapping analyst coverage, which results in stronger predictability than using industry classifications that could be too broad, too narrow, or poorly defined for a given stock.
Moreover, using analyst coverage results in clusters of peer stocks in which the pool of analysts behave similarly. Therefore, earnings surprises may be expected to be more similar. In a multiple regression explaining current surprises using past peer surprises, a peer surprise feature defined by analyst coverage dominates one defined by industry peers.
3: Pay Attention to Announcement Timing, Guidance, and Variance Across Estimates
There is academic evidence that companies who tend to miss their numbers are late to announce, or they announce on days with high news volumes, near weekends or holidays. Companies for which analysts are uncertain (as evidenced by high cross-sectional variance in estimates) also tend to miss their targets.
When a miss is expected, management often seeks to reduce the share price impact of weaker-than-anticipated results, and sometimes attempts to “take a bath” by engineering future quarters’ losses into the current quarter.
These effects and management behavior can be captured with variables such as management guidance, historical beats and misses versus guidance, the nominal and relative timing of announcements, expectations, and variance across analysts’ estimates.
Why These Approaches Work
By looking at the average daily cross-sectional Spearman rank correlation below, we observe that each group of features – which we categorize as related to Experts, Trends, or Management – is not very correlated to the others.
Insight/2023/03.2023/03.01.2023_3%20Information%20Sources%20That%20Can%20Help%20Improve%20Earnings%20Forecasts/01-average-daily-cross-sectional-spearman-rank-correlation.png?ssl=1){kind=icon}
We can also measure the accuracy of forecasts using a hit rate: When the forecasted surprise is above or below a minimum threshold, what percentage of the time is the forecast directionally accurate in predicting beats and misses? Below we measure the hit rates for each component and the overall forecast.
Insight/2023/03.2023/03.01.2023_3%20Information%20Sources%20That%20Can%20Help%20Improve%20Earnings%20Forecasts/02-eps-and-revenues-forecast-accuracy-by-hit-rate.png?resize=646%2C201&ssl=1)
To measure a composite-surprise prediction for stock returns, we backtested a simple long/short, top/bottom decile portfolio based on the TrueBeat value, rebalanced daily and equally weighted. Below we show the non-compounded cumulative return since 2000.
Insight/2023/03.2023/03.01.2023_3%20Information%20Sources%20That%20Can%20Help%20Improve%20Earnings%20Forecasts/03-backtest-fq1-eps-truebeats.png?ssl=1)
The signal’s alpha is robust and is confirmed by its exposure to sectors. Returns attribution is confirmed by market beta, factors, and industries, for example.
In Summary
The recent explosion in data about publicly traded companies, and the concurrent growth in analytical techniques such as machine learning, have opened many opportunities for quants and data scientists to improve their predictions of returns.
There’s an embarrassment of riches when it comes to data and quant work, with much of it in large and often unstructured datasets. But perhaps lost in the flood of data and tools is the fact that one can take structured datasets with long histories and use them to improve one’s predictions. And not just with returns, but with more fundamental measures such as EPS – all through careful feature engineering with an eye towards the fundamental, behavioral, and economic rationales behind each feature.
A well-designed forecast of earnings built in this way should, in most market conditions, translate to an effective prediction of the cross section of returns.
Use our TrueBeats Model to accomplish all of the above – it combines the above research findings into an explicit prediction of earnings and sales surprises. Contact us for a free trial, and while you’re on the page, feel free to request information on some of our other datasets and signals.
