Astute investors know that fundamental research, along with the expertise and time needed to do it right, is essential for stock forecasting.<\/p>\n\n\n\n
With advancement in NLP and machine learning science, one could only assume how transformational automatically parsing and analyzing corporate filings would be. Greater accuracy, wider corporate coverage, faster time-to-market, and lower cost are just some of the benefits.<\/p>\n\n\n\n
Therefore, it\u2019s not surprising to learn that many have invested significantly in R&D to crack the footnotes code. The main challenges in automating analysis of corporate earnings reports are the complexities of many disclosures and the lack of standard reporting format, particularly for one-off, transitory gains\/losses hidden<\/a> in footnotes of reported and consensus earnings.<\/p>\n\n\n\n Recently, we\u2019ve partnered with New Constructs (NC), an 18-year veteran in independent research. <\/p>\n\n\n\n Our research focused on validating the predictive nature of NC data by comparing consensus earnings to the \u201ctrue\u201d Core Earnings. Our goal was to identify the stocks where the street gets earnings most wrong<\/strong>.<\/p>\n\n\n\n Our Hypothesis: <\/strong>A buy long<\/strong> signal is triggered when a company\u2019s true earnings (Core Earnings) is understated compared to consensus, i.e., Earnings Distortion is negative. When such an understatement<\/strong> is not understood by the market, the street under-estimates future earnings, and thus provides an opportunity for a buy long <\/strong>entry. Conversely, a sell short<\/strong> signal is triggered when a company\u2019s true earnings (Core Earnings) is overstated compared to consensus, i.e., Earnings Distortion is positive. When such an overstatement<\/strong> is not understood by the market, it provides opportunity for a sell short<\/strong> entry.<\/p>\n\n\n\n Images 1 & 2: Reported earnings is understated presents a buying opportunity and conversely when reported earnings is overstated it presents a selling opportunity.<\/em> Our research focused first on validating New Constructs\u2019 data by identifying companies that respond similarly to earnings distortions. The graph below demonstrates clustering, an unsupervised<\/a> machine learning technique, geared to group stocks by how similarly their stock price responds to certain Earnings Distortions. In contrast to grouping assets by predetermined GICS code (Global Industry Classification Standard) or industry sectors, we let the data determine how to group assets based on certain similarities, specifically, how they respond to market changes (such as in volatility or new earnings distortion).<\/p>\n\n\n\n Image 3: Clustering used to group similarly behaving stocks post-earnings having similar Earnings Distortions. A spring-force graph is used to demonstrates elements that attract based on similarities and repel based on dissimilarities.<\/em> Since earnings are released just once per quarter, we looked to evaluate how long past earnings release dates it takes for stock prices to adjust to Earnings Distortion (also called measuring alpha decay). In other words, we looked to assess how long on average it takes for the market to adjust a stock\u2019s price displacement due to Earnings Distortions.<\/p>\n\n\n\n We first looked at long investment opportunities resulting from negative Earnings Distortions. We divided all assets in the Russell 1,000 into deciles (or 10 bins) each representing Earnings Distortion as a percentage of GAAP earnings.<\/p>\n\n\n\n Bin 1:<\/strong> Decile with the most negative distortions.<\/p>\n\n\n\n Bin 10:<\/strong> Decile with least negative or positive distortions.<\/p>\n\n\n\n The image below showcases how larger distortions affects the price of stocks more meaningfully and for a longer time frame (longer alpha decay, compared to lower distortions.<\/p>\n\n\n\n Image 4: Comparing lowest decile in orange to highest decile in blue provides clear evidence that the higher the distortion the more pronounced the subsequent price action.<\/em> Given the strong evidence of predictive information in New Constructs\u2019 data, we moved swiftly into machine learning models in order to conduct backtests on whether earnings distortions, as a feature, can be fed into a machine learning classifier. (A classifier is a machine learning objective which targets a predetermined state.) In our case, we looked for one of three states: buy<\/strong>, sell<\/strong>, or do nothing<\/strong>. Our goal was to identify when to buy a stock, how long to hold it, and when to ultimately sell it in order to maximize the likelihood of profit.<\/p>\n\n\n\n This is the phase in which we apply the heaviest machine learning research. It involves model training, hyper-parameter tuning, cross validation, back testing, perpetual roll-forward testing, and quite a few additional AI disciplines.<\/p>\n\n\n\n Our model\u2019s objectives were to forecast a company\u2019s future core earnings<\/strong> and compare it to the street\u2019s projected earnings<\/strong>. In addition, since we are looking to compare companies of various sizes, we needed to normalize earnings distortion to a company\u2019s size. We chose to use market cap for this particular task.<\/p>\n\n\n\n Image 5: Normalizing distortion values as percent of the underlying company\u2019s market cap.<\/em> In summary, our goal was to take New Constructs\u2019 raw data and form machine learning models that will forecast a company\u2019s true future earnings (after adjusting for one-off, non-recurring distortions), compare such future earnings to the street\u2019s estimated earnings and, when a meaningful variance is identified, act by either buying or selling the underlying stock.<\/p>\n\n\n\n The chart below demonstrates our process in taking New Constructs\u2019 data and conditioning it for a model portfolio investment as follows:<\/p>\n\n\n\n \u25cf Breaking the constituent universe (Russell 1K) into quintiles by market cap.<\/p>\n\n\n\n \u25cf Applying clustering to further hone in on similarly behaving stocks to earnings distortions.<\/p>\n\n\n\n \u25cf Feeding earning distortions factors into machine learning classifiers.<\/p>\n\n\n\n \u25cf Generating buy\/sell\/do-nothing signals.<\/p>\n\n\n\n \u25cf Constructing the portfolio by identifying:<\/p>\n\n\n\n Image 6: Demonstrating the life cycle of utilizing New Constructs\u2019 data to form market cap quintiles, on which we apply clustering to fairly comparable constituents\u2019 distortions. Subsequently, we train an ensemble of models to vote a \u201cbuy,\u201d \u201csell,\u201d or \u201cdo nothing\u201d classification.<\/em> The Results<\/strong><\/p>\n\n\n\n The results below are of a backtest simulation on Lucena\u2019s DAS (Data Analytics Services) platform. We were able to empirically validate how actionable New Constructs data is for both long and short investment scenarios. To further validate our models and ensure no look-forward bias, we are now perpetually paper trading the models into the future.<\/p>\n\n\n\n Backtest \u2013 New Constructs Long Only \u2013 View Backtest<\/a><\/p>\n\n\n\n Image 7: New Constructs Long Only Earnings Distortion Portfolio.<\/em> The backtest screens the Russell 1000 daily for the stocks with the most meaningful variance between the street\u2019s and our model\u2019s core earnings estimate. The portfolio\u2019s goal is to outperform the Russell 1000 equally weighted benchmark in risk adjusted return as measured by its Sharpe ratio. In addition, we conducted a backtest to identify short signals based on an overstated street outlook as compared to a company\u2019s true earnings outlook (core earnings).<\/p>\n\n\n\n Backtest \u2013 New Constructs Short Only \u2013 View Backtest<\/a><\/p>\n\n\n\n Image 8: New Constructs Short Only Earnings Distortion Portfolio<\/em> The backtest screens the Russell 1000 daily for the stocks with the highest earnings distortions. The portfolio\u2019s goal is to outperform SH (inverse S&P 500 ETF) in risk- adjusted return as measured by its Sharpe ratio.<\/p>\n\n\n\n Data is the fuel that feeds the AI engine. When a proprietary data source such as New Constructs presents insights into a company\u2019s true Core Earnings before the street realizes it, investment opportunities arise. <\/p>\n\n\n\n Next week, I will be covering another exciting set of strategies that are based on social media sentiment, and specifically, crowd voting sentiment. We will learn about leveraging cutting-edge machine learning technology to identify superusers and score confidence at the aggregate towards specific investable assets. For reference see Pynk.io<\/a><\/p>\n\n\n\n If you are a data provider with unique data that could be useful for investment, we want to talk to you. In addition, if you are an investment professional looking for winning investment portfolios, feel free to reach out to us. We\u2019re happy to grant you trial access to a model portfolio that suits your investment style and mandate.<\/p>\n\n\n\n Have a great week!<\/p>\n\n\n\n![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-1.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\nThe Research:<\/strong><\/h2>\n\n\n\n
Step 1 \u2013 <\/strong>Validating the Raw Data<\/h3>\n\n\n\n
![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-2.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\nStep 2 \u2013 Assessing Alpha Decay<\/h3>\n\n\n\n
![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-3.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\n Step 3: Portfolio Construction:<\/h3>\n\n\n\n
![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-4.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\n![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-5.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\n![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-6.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\n![\"\"](\"\/campus\/wp-content\/uploads\/sites\/2\/2022\/02\/neuravest-algo-trading-7.png\")
<\/figure>\n\n\n\n
Source: Neuravest Research Platform DAS (Data Analytics Suite)<\/em><\/p>\n\n\n\nConclusion<\/strong><\/h2>\n\n\n\n