Unlock the Power of Quantitative Strategies: Explore Our Cutting-Edge Website Today!


AI case study: Long/Short Strategy

Javier Cárdenas


No Comments

At ETS Asset Management Factory we are constantly improving our strategies by promoting innovation based on our data, algorithms, team and experience. Each of these elements serves a distinct purpose, collectively forming our company’s four main pillars.

Following this philosophy, in today’s post we will be using an advanced algorithm to improve a module of the Alternative Data-Driven Investment (ADDI) strategy developed by ETS Asset Management Factory, which is an automatic Long – Short investment strategy that aims to obtain stable performance de-correlated from the market and with a limited drawdown risk.

The algorithm that we have developed is a tailored deep neural network and we will be using it to improve the risk associated with our Long-Short strategy. This comes after our recent posts, where we have been looking into key aspects of neural networks, their applications, and various case studies demonstrating their relevance in the field of finance (see for example 1, 2 or 3).

Why using neural networks?


For quantitative investors, acquiring data marks the halfway point in their journey. But one of the most important and intriguing phases lies ahead, where an endless number possibilities unfolds: How are you going to transform this data into signals?

You can opt for traditional statistical methods to scrutinize your hypotheses, or you may venture into the realm of advanced algorithms like machine learning and deep learning. Perhaps your fascination with several macroeconomic theories makes you want to investigate its applicability in the foreign exchange (FX) market. Or maybe your passion for comprehending the fundamentals of the companies may lead you toward the path of quantitative value investments. Each of these research paths is not only valid but also interesting for exploration.

Methodology at ETS

In our case, there’s a single guiding principle when it comes to choosing a research direction: Embrace innovation.

A pile of rocks ceases to be a rock pile when somebody contemplates it
with the idea of a cathedral in mind.

Antoine de Saint-Exupéry

The rationale behind this is quite straightforward—if we don’t innovate, we won’t stand out, and our chances of success would diminish. So, whether we’re developing a fresh strategy, whether it relies on classical statistics or involves extracting insights from a company’s financial statements, we always try to approach it with novel and innovative methods. We create concrete tests for specific scenarios, detect anomalies in financial statements or adapt our models to specifics problems.


Neural networks can be particularly appealing to solve specific problems due to their adaptability and flexibility, which helps us in the process of developing innovative techniques. Once again, your creativity knows no bounds here. Would you like to use your financial data as time series using Transformers? Perhaps engage in continuous learning through reinforcement learning? Or maybe your goal is to devise novel structures that better address your unique problem?

As you can observe, there are numerous ways to employ these techniques. However, it’s essential to exercise caution; there’s no magic formula here. Just like with any endeavor, we should always begin with the basics, and sometimes, a simple linear regression can prove remarkably effective 😉 .

Returning to today’s focus, we’ll be using the power of a deep neural network to forecast a company’s risk using as inputs data from the financial statements and historical prices.

Our model

That being said, we will assess the performance of our model by comparing it to simpler methods like historical volatility.

Before we dive in, we’ll assume you are already familiar with some key concepts of neural networks and how we were using them to derive the distribution of our predicted output. You can find more information on this topic in our previous posts here and here.

Furthermore, although we would love to explain in detail all the processes that led to the development of our model (the specific data selected, the training period, the structure of the deep neural network, the training loss, etc..) for today’s purpose we will focus our attention only on the improvements against our benchmark and the results obtained after using it in our investment strategy.

So continuing the later post of measuring uncertainty in time series data we will try to estimate the risk of the company by predicting the quantiles of the expected price return to different time horizons in the future, in our case from 5 days up to 90 days ahead.

Below, you can find an example of how the quantile predictions for different time horizons in the future (in blue) would appear after the model has been trained. The wider the quantile intervals predicted the greater the risk of our investment. In purple is what actually happened after making the predictions.

Evaluating our model

As discussed earlier, before using our model in our strategies, we will compare its predictions with the ones obtained using a simple transformation of past volatility. Is a simple transformation of past volatility better than a complex algorithm?

To evaluate both our model and our benchmark we have compared their quantile predictions with the observed returns. As an example, among all our predictions for the quantile 0.9 we would expect to see, on average, the prices return below that quantile predicted 90% of the time.

And this is what we try to evaluate in the graph below (all results shown come from the test set). In the left plot we can compare the theoretical and real % coverage. On the plot of the right we see the difference of these coverages (theoretical – real) which we have called coverage error. As an example, for the quantile 0.2, there is a coverage error of almost 0.4% meaning that, on average, we have observed the 20.4% of our data below those values instead of the theoretical 20%.

We have averaged all the quantiles coverage errors by projection window (5, 10, … days) and we compare the results obtained by our benchmark and our AI model. Below we can see that our deep learning model performs better (fewer average % coverage error) and we are ready to introduce our model in our strategy.

AI inclusion in ADDI

ADDI is a beta neutral leverage equity portfolio (beta ~ 0.1) capable of generating alpha both in bearish and bullish market regimes, with a limited net exposure to the market and a reduced risk profile.

The long leg of the strategy selects stocks of high quality and with a low volatility bias. Therefore, the stock risk estimation is an important task in the process. In the short leg the stock risk estimation is also an important calculation to take care of, as the strategy tries to avoid those stocks with extreme risk or with very low risk.

We can measure the stocks risk by the stocks historical volatility with different calculation periods in the long and short parts of the strategy.

In order to improve ADDI risk analysis, we are going to test the inclusion of the deep neural network algorithm shown before to replace current risk calculation process.


Testing the new deep learning model on the Long – Short strategy investing over the S&P 900 components you can see that the results both in terms of performance and risk get better:

New deep learning model on the long-Short investment strategy's results
  • Total return shows higher figures than the original version
  • Volatility is reduced
  • Sharpe Ratio Improves
  • Risk diminishes in terms of drawdown and VaR
  • Maximum 1 year rolling runup is higher.


In this post we have shown an example of an advanced algorithm model used to improve a Long-Short quantitative strategy over equities (ADDI), developed by ETS Asset Management Factory. We have presented how neural networks can be harnessed to improve and manage more accurately specific tasks in quantitative investment products, thus improving final results.

However, the applicability of the model doesn’t have to stop here; we can use the algorithm for various other strategies. For instance, we can employ it to select companies with the highest Sharpe ratio, those with a positive skew, or even for implementing a pairs trading strategy. Can you think of any other strategies?

Thanks for reading and have a look at our web for more!

Inline Feedbacks
View all comments