JP Morgan's Take on Valuation Strategies based on Machine Learning and RavenPack Sentiment

J.P. Morgan provides ideas on implementing Machine Learning algorithms to stock selection processes based on valuation. They show how RavenPack data is a useful overlay to their strategy.

Following the popularity of their recent report, "Big Data and AI Strategies", JP Morgan’s Global Quantitative & Derivatives Strategy team [JPM] has turned their attention to stock selection - using Machine Learning to come up with a valuation based strategy. In their new report, they attempt to predict the “fair value” of stocks using a large number of equity characteristics, and quantify a “mispricing” signal that either buys or shorts stocks that are under- or overvalued.

So why is this interesting to RavenPack? Well, the report shows that including a news sentiment overlay based on our data improves the overall performance of their mispricing signal - yielding an Information Ratio of 1.55 (and 1.85 without transaction costs).

Below, we provide an overview of the 5-step approach considered by JPM when constructing their valuation strategy. The backtesting universe consists of about 2000 stocks across 39 countries - all belonging to the MSCI AC World index. For a more detailed description of their universe, transaction cost assumptions, and general methodology, we refer to the full report.

Let’s get started….

Step 1- Feature Construction

As part of creating a mispricing signal, JPM considers over 37 stock characteristics that could potentially help to predict the price-to-book of a stock, with perspectives ranging from valuations, profitability, operational efficiency, quality, growth, sentiment, and risk. Table 1 lists the features that are included in their model.

Feature Construction- JP Morgan 37 stock characteristics

Step 2 - Machine Learning Models

Taking the above features as input, JPM fits a linear and elastic net regression as well as a boosted tree model (XGBoost). They find that the latter model provides the strongest performance with an IR of 0.89. A combination of all three models does similarly well with an IR of 0.88 (see Figure 18).

JPM fits a linear and elastic net regression as well as a boosted tree model (XGBoost)

Step 3 - Gross Profit Overlay

Intuitively, investors are interested in buying stocks at a reasonable price (i.e. value stocks) which also demonstrate their ability to make profits. By combining the mispricing model with Gross Profit to Assets (GP/A), JPM is able to improve strategy returns and increase the IR from 0.88 to 1.26 - also outperforming a combination strategy based on ROE as an alternative measure of Profit (see Figure 20).

By combining the mispricing model with Gross Profit to Assets (GP/A), JP Morgan is able to improve strategy returns and increase the IR from 0.88 to 1.26

Step 4 - Combining vs. Filtering

In the above “combinative” approach, investors may end up with stocks that are either largely undervalued (but not really profitable), or with a very high profitability (but not necessarily undervalued). Instead, JPM tries to play the “mispricing” signal by first looking at stocks that are highly undervalued or overvalued, and then using the GP/A to filter stocks that they want to retain in the long and short portfolios - only keeping the top or bottom 40% based on GP/A, respectively. This approach brings further IR improvements, increasing it from 1.26 to 1.46, primarily driven by greater returns (see Figure 21).

JPM tries to play the “mispricing” signal by first looking at stocks that are highly undervalued or overvalued, and then using the GP/A to filter stocks

Step 5 - Including News Sentiment

As a final step, JPM shows how introducing RavenPack’s news sentiment signal as an overlay provides further improvements to their strategy. In particular, removing stocks with negative sentiment in the long basket (i.e. undervalued stocks with high profitability) and positive sentiment in the short portfolio helps to improve risk-adjusted returns. Figure 30 shows the results of their strategy, where the Information Ratio (after transaction costs) is increased from 1.46 to 1.55. This is mainly due to an increase in returns, while volatility is almost the same. Without transaction costs, this strategy would give an IR of 1.85.

introducing RavenPack’s news sentiment signal as an overlay provides further improvements to their strategy

Note that Marko Kolanovic and Rajesh T. Krishnamachari, the Global Head and Vice President of the Quantitative and Derivatives Strategy at J.P. Morgan, respectively, spoke at the RavenPack 5th Annual Research Symposium: The Big Data & Machine Learning Revolution, that took place September 19th in NYC.

You can request the full report here.

Discover the RavenPack Platform

We offer more than 90 dashboards, built with over 220 curated datasets and powerful visualisation tools.

DISCOVER THEM HERE