Earnings Surprises, Insider Info, Logistic Regression: Insight Into Post-Earnings Trading

by: Ramin Mohammadi


Many financial news sources place great importance on earnings releases.

While important, these sensationalized stories do a poor job explaining how a stock trades post-earnings.

In order to further understand why a stock moves after an earnings release, I performed an econometric analysis in the form of a logistic regression.

With summer earnings releases right around the corner, many traders find themselves asking a similar question: Is it possible to consistently profit during earnings season? Of course, if you were to ask around Wall Street you would turn up a variety of answers, ranging from obtaining insider information, to trading breakouts, or even trading options straddles. However, all of these strategies have their flaws.

Obviously, insider trading is illegal (as well as pretty much fruitless, as we will see later). Breakouts can be false, are difficult to time and will throwback a large percentage of the time, whipsawing day-traders into losses more often than gains. Options spreads become heinously expensive around earnings, due to high implied volatility, eating up profits with high theta/vega. All those things considered, if there were in fact a fundamental data point, released during an earnings report, that could correlate to post-release trading activity, maybe speculators would then have a starting point from which to place their bets.

Recent Nobel laureate Eugene Fama and a significant percentage of academia would argue that an equity's price reflects all possible public information at any given moment (EMH). If that is the case, one could hypothetically profit by correctly outguessing the consensus analyst estimates, which get published weeks before. If you think you might have greater insight to the true values of these figures, you could theoretically place strategic trades to capture the surprise. By no means am I arguing a case for insider trading; in fact, any truly savvy trader could ascertain that the payoffs are nowhere near the risk involved. That said, could insider information consistently produce the "sure thing" that every risk-taker strives to find? The following model might shed some light on this situation.

The Model

Qualitative binary response models have been used since the early 1940s as a means to assess the probability of two outcomes, given a set of independent variables. Mostly used by economists for policy related issues, these binary response models can provide insight into how probabilities can be affected by exogenous factors. In terms of financial data, the most appropriate model would be a logistic regression, mostly because of the leptokurtic nature of the probability space associated with the logistic cumulative distribution function.

Examining the past six earnings reports for 30 companies, diversified by market capitalization and industry, can explain how certain surprise percentages for four key fundamental values impact whether a stock closes up or down in the next trading session. With any model there are critical assumptions that underlie the validity of the experiment in question. These assumptions will be discussed later; however, here are some concerns that can be immediately quelled.

Critics might point out that surprise might not have an immediate effect on a stock's price and that a longer time frame must be considered in order to properly price the equity in question. This model does not intend to capture that effect. Instead, it seeks to explain any immediate returns obtainable only through insider knowledge of a company's press release. Others may point out that 180 data points might not be enough to obtain strong levels of statistical significance. For that reason all standard errors in the study were obtained via a parametric bootstrap, a method that uses simulation, in this case one hundred thousand random draws, to properly capture any kurtosis (but not skew in this case) in the probability space of the standard errors, resulting in far superior results than any other econometric technique.

Lastly, there are those who may say that every company trades differently, what might work for one won't for another. That point is well-taken and possibly correct. However, this study tries to capture a relationship in a universal context, to see if earnings can correlate to specific trading action.

The Assumptions

Following the framework of any econometric/quantitative model, a list of assumptions has to be made, given the general simplification of the most complex network of supply and demand in today's day and age. Starting from the top, we must assume that any insider, if they do exist, has not affected the securities price significantly enough to change post-earnings expectations. If this were true then the equity's price would already reflect post-earnings data, making any attempt at modeling or profiting from earnings trades futile. The rest of the assumptions involve the concept of endogeneity and really serve as the backbone for any model in applied statistics.

Essentially, a model is said to be endogenous when there is correlation between the independent variable and the stochastic error term. In this particular model structure, it can result from simultaneity between the dependent and independent variable or omitted variable bias. Given the nature of this website, the CAPM and alpha in general, it seems acceptable to eliminate the possibility of endogeneity (strictly in this framework).

The Data

The data encompasses 30 companies, with seven independent variables, three of them controls. To maintain the concept that everyone has the same starting information, the earnings figures are surprise values (in percentages), that capture the difference between the mean analyst estimate and the actual figure. One could argue many relevant metrics, but revenue surprise, EPS surprise, EBITDA surprise and cash flow per share surprise seemed sufficient for this purpose.

The controls, which are dummy variables, serve as a filter for any background noise that could arise from the differences in the individual earnings dates. This includes the day of the week, whether said week is an expiration week for monthly options and whether the company reported earnings before or after market hours. The dependent variable is a binary value indicating whether the stock closed up or down in the next full trading session. The model itself was coded into STATA, a statistical package used frequently by economists around the world.

The Results

The model returned some pretty interesting results that may surprise some traditional investors.

Click to enlarge image.

Source: Thomson Reuters/Matoaka Capital LLC.

Clearly visible in the data output, the only statistically significant variable is whether or not earnings occurs on an options expiry week. Somewhat significant (not necessarily from a statisticians point of view), we have the release time. The most relevant variables include those whose coefficients have non-zeros in the first two decimal spaces, as their effect would have a substantial impact. In terms of the coefficients, the fundamental variables would be interpreted as a one percent increase in said value would represent the corresponding coefficient increase/decrease in the probability of the stock closing up. The dummy variables represent the effect on probability if categorized as a yes versus a no.

The Interpretation

The model results yielded very insignificant results, implying a negligible relationship between earnings surprise and post-earnings trading action. The one significant variable was completely exogenous of the stock itself, instead relating to the general derivatives market. All things considered, the intent of this discussion is not to disprove the basis of fundamental valuation, but more so to try and point out that immediately speculating on the direction of a stock post-earnings is a fruitless endeavor, with a negative expected value.

In the long term, these earnings values may or may not impact the price of the stock; however, in the short term, you are better off flipping a coin. That said, certain important factors were purposely left out of the model. An interesting study would be to relate the magnitude of returns, regardless of direction, to the magnitude of earnings surprise, to possibly gain some insight there. Also, a metric for market sentiment, such as the VIX or the consumer confidence index could prove useful in gauging the extremity of price movements.

In conclusion, certain individuals will always try and tell you that they have a profitable earnings season strategy. Seasoned options "risk doctors" can definitely make money through an active, gamma scalping trading approach, so there are definitely others who can as well. For the most part, however, these "earnings gurus" are not being truthful and should be disregarded.

Disclosure: The author has no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours. The author wrote this article themselves, and it expresses their own opinions. The author is not receiving compensation for it (other than from Seeking Alpha). The author has no business relationship with any company whose stock is mentioned in this article.

Additional disclosure: The information contained herein is not necessarily complete and its accuracy is not guaranteed by Matoaka Capital L.L.C., its operating entity or the principals therein. Principals of Matoaka Capital L.L.C. may or may not hold or be short securities discussed herein, or any other securities, at any time. If you are interested in the raw data/STATA code, please message me directly.