Forecasting Earning Surprises with Machine LearningHow to predict which companies will Beat or Miss their Analyst Earnings EstimatesCharles BrecqueBlockedUnblockFollowFollowingFeb 12Listed companies produce quarterly earnings reports which can cause significant price movements when the results deviate from what the analysts had estimated.
This is because according to the Efficient-market hypothesis, asset prices fully reflect all available information and will as a result factor in consensus estimates.
In this article we are going to see how we can use Machine Learning to predict whether a company will beat or miss its estimates.
The DataWe consider EPS analyst estimates from the Thomson Reuteurs I/B/E/S Estimates database and was downloaded from Sentieo.
The database gathers and compiles the estimates made by analysts for more than 20 measures.
For each company, we are given the Mean, #Estimates, Low, High and Actual values of the estimates as shown bellow:Unfortunately, with this database we only have 70 data points per company, which isn’t enough to predict the earnings of one company based on previously announced results and their Beat/Miss vs estimates, but we can instead reframe the problem to increase the number of data points.
Instead of asking ourselves whether a company will beat or miss the estimates, we can ask whether the estimates will be higher or lower than the actual values.
We will then normalise the values in order to aggregate them.
In this case the features we will be considering for our model are:# EstimatesLow/Mean %High/Mean %Actual/Mean %We then decided to aggregate the estimates by sector in order to test the hypothesis that the analysts (in)ability to forecast earnings accurately would be tied to the nature of the firms.
For this study we are going to focus on healthcare stocks.
We then took over 6000estimates for the following 117 companies:‘AAC’, ‘ABT’, ‘ABBV’, ‘ACHC’, ‘XLRN’ ,’ACOR’,’AERI’,’AGIO’,’AIMT’,’AKCA’,’AKBA’ ,’AKRX’,’ALXN’,’ALGN’,’ALKS’,’AGN’,’ALNY’,’AMRN’,’AMGN’ ,’FOLD’,’ARRY’,’ASND’,’AZN’,’ATRA’,’ATRC’,’AVNS’ ,’BHC’,’BAX’,’BDX’,’BCRX’,’BMRN’,’TECH’,’BEAT’,’BLUE’ ,’BSX’,’BMY’,’CBM’,’CAH’,’CSII’,’CELG’,’CNC’,’CRL’,’CHE’,’CBPO’ ,’CNMD’,’CORT’,’CRY’,’DVA’,’XRAY’,’DXCM’,’EHC’,’ESPR’,’EXAS’ ,’EXEL’,’FGEN’,’FMS’,’GHDX’,’GILD’,’GMED’,’GRFS’,’HAE’,’HALO’,’HCA’ ,’HCSG’,’HSIC’,’HLF’,’HRTX’,’HRC’,’HZNP’,’HUM’,’ICUI’,’IDXX’,’IMMU’ ,’INCY’,’INVA’,’INGN’,’INSM’,’IART’,’ISRG’,’IONS’,’IOVA’,’JAZZ’ ,’LGND’,’LIVN’,’LMNX’,’MGLN’,’MASI’,’MCK’,’MMSI’,’MOH’,’MSA’,’MYL’, ’NEOG’,’NEO’,’NBIX’,’NVO’,’NUS’,’NUVA’,’OPK’,’OFIX’,’PDCO’,’PAHC’,’DGX’, ’RMD’, ’SEM’,’ONCE’,’STE’ ,’SYNH’,’TFX’,’COO’,’USPH’,’UNH’,’UHS’,’VAR’,’VRTX’,’WBA’,’ZBH’Processing the DataWe uploaded the data to AuDaS, a Data Science and education platform built for Analysts by Mind Foundry.
In order to improve the accuracy of the model, we created a new column which represented whether the actual value was higher (1) or lower (-1) than the actual value (as opposed to a %).
We can also visualise the data through the automatically generated histograms and see how the Beat/Misses distribute cross the other features.
Building our Machine Learning modelAs we are predicting the Beat/Miss column we will build a classifier whilst excluding the Actual/Mean column.
We will also stick to the recommended training configurations which are automatically chosen by AuDaS to prevent over-fitting (10 fold cross validation and 10% hold-out for validation purposes).
AuDaS will then start searching for the best Machine Learning model using Mind Foundry’s proprietary optimiser, OPTaaS, which has quickly established itself as the Quant fund industry’s favourite tool for fast global optimisation and hyper-parameter tuning.
In less than a minute, AuDaS had already tried 37 different Machine Learning models and the best found solution was a simple Gradient Boosting Classifier.
Hovering over the model reveals its parameter values:The relative feature relevance for this model suggests that the Low/Mean, High/Mean ratios contain the most information.
Testing the ModelWe can then test the model on the 10% hold-out:AuDaS achieved a classification accuracy of 69.
4% and the final Model Health Advice was good.
Conclusion and extensionsWith relatively few features, AuDaS was able to build an accurate forecasting model that can support Investment Analysts in their review of IBES estimates.
This can enable them to anticipate significant price movements.
An extension to this study would be to use AuDaS’ clustering feature to group estimates across multiple sectors.
This would enable us to test the hypothesis that other factors such as corporate governance will impact the Beat/Miss ratio of estimates more than the sector.
If you are interested in finding out how our Quant and Fundamental Hedge Fund customers use AuDaS for their Risk analysis and Investment Management please don’t hesitate to reach out to me by email or LinkedIn.
You can also read some more case studies bellow:Risk Management with ClusteringHow to uncover structure bellow the surfacetowardsdatascience.
comValue Investing with Machine LearningYour favourite holding period doesn’t have to be forever…towardsdatascience.
comAugmenting Investment Analysts with Data ScienceHow fundamental investing can benefit from Machine Learningtowardsdatascience.
comTeam and ResourcesMind Foundry is an Oxford University spin-out founded by Professors Stephen Roberts and Michael Osborne who have 35 person years in data analytics.
The Mind Foundry team is composed of over 30 world class Machine Learning researchers and elite software engineers, many former post-docs from the University of Oxford.
Moreover, Mind Foundry has a privileged access to over 30 Oxford University Machine Learning PhDs through its spin-out status.
Mind Foundry is a portfolio company of the University of Oxford and its investors include Oxford Sciences Innovation, the Oxford Technology and Innovations Fund, the University of Oxford Innovation Fund and Parkwalk Advisors.
.