Quoniam AM turns to machine learning for non-linear stock relationships

Cause and effect in the financial markets are not always easy to understand, especially when the relationship can’t be plotted as a straight line on a graph, and changes are difficult to predict. Frankfurt-based Quoniam, which has more than $30 billion in assets under management, is using machine learning to forecast returns where there are non-linear relationships between stock characteristics and the returns on these assets.

Volker Flögel, the firm’s head of research, says relationships in financial markets are often not directly connected. For example, it would seem obvious that stocks that yield low dividends perform worse than those that yield high dividends. However, companies that do not pay dividends are not necessarily performing poorly, especially if they are growing quickly. 

“A linear relationship that assumes that a high dividend yield is better than a low one cannot reflect this [growth or these returns],” Flögel says. Quoniam augments its linear predictive models with machine learning when forecasting returns.

To model the relationships, the asset manager uses gradient boosted trees, a machine learning technique for optimizing the predictive value of a model using an ensemble of decision trees. “This class of models is well suited for data with a low signal-to-noise ratio, which is what we typically observe in financial markets,” he says.

Quoniam buys raw data from partnering companies to use within these models, performing its own data normalizations and cleansing. The firm relies on open-source software for basic machine learning algorithms, though it does not use complete off-the-shelf solutions, as these tools don’t provide the firm with enough flexibility and transparency, Flögel says.

“We want to have full control over how the algorithms are applied to the data,” he says. “So the pipeline, data preparation and parameterization, and so on, is all done in-house. However, where available, we will always use packages that are tested and where a large set of algorithms are already implemented.”  

Such software packages are programmed and developed during open-source collaboration and focus on areas including time series analysis, visualization, or machine learning. Flögel says the packages have to be applied by the user, and if something is missing within them, they have to edit it themselves.

“We have a basic programming language like R or Python, and then there are packages that are developed by the open-source community that contain different functions depending on the purpose of the package. It can be a package for visualization or it can be a package for machine learning,” he says.   

Flögel says the information Quoniam looks at is less focused on studying short-term measures, such as current order book data, which would interest a quant team. Instead, the firm looks at information such as the value of a company, what corporate insiders are doing in their own stocks, as well as analyst and sentiment information about what other market participants are thinking.

“We are not running any intraday models and positions, but we are trying to forecast returns for a holding period of a year or even a little bit longer,” he says.  

Calculating the actual return forecast using new data can be done in seconds, but Flögel says training the models is a longer process. For example, in a regression model, calculating coefficients for a forecast would require having all the historical data of the stock. In estimating the regression, the model would need to have different types of weights for factors such as value, quality, or sentiment. To calculate the forecast, the weights would be left constant, perhaps being re-calculated once a month. 

“If you want to calculate the forecast daily, you have just one observation for each stock, and that is very easy to calculate. It is just a matter of seconds,” Flögel says. 

Quoniam’s machine learning models are built within its research department, which includes analysts with a diverse range of backgrounds such as IT, business, finance, and accounting. The research team provides forecasts in terms of risks and return forecasts for different assets to their colleagues in portfolio management.

The portfolio management team takes the forecasts and applies optimization techniques. Finally, the portfolios are implemented by the trading desk.  

Next stop: fixed income

Over the next year, Quoniam plans to start working on applying machine learning techniques to other asset classes.

It currently runs models for forecasting fixed income, but they don’t capture non-linear relationships. Flögel says applying machine learning techniques, as it has done for equities, could improve the accuracy of these forecasts.   

“We currently use machine learning for our equity models. But we have to also run fixed-income models, and obviously, that is something where we want to learn to apply machine learning techniques as well,” he says.  

Applying these approaches to other asset classes could also be a check for robustness to see if the methodology works in a different setting, he adds. 

Quoniam is also working on further developments—for example, leveraging its models to analyze even larger amounts of data. “All this is happening against the background of establishing machine learning as a standard method in our research process,” Flögel says. “This will increase the efficiency of our research process because we will be able to carry out research projects faster than before.”