TECHNOLOGY

At Axyon AI we focus on developing the best deep learning solutions to boost clients' investment strategies.
Hence, our very first step is to understand clients' needs in regards to the investment objective like the investment universe, the investment horizon, the benchmark and the risk management constraints and anything else that needs to be considered.

Our database is continuously updated with new high-quality predictive and alternative sources of data.

Why is it important?
This step is important to ensure a structured data research process as new, and original data sources can improve alpha generation thanks to our AI models*.

*AI modelling is the creation of a logical decision-making system based on a certain set of available data.

Data is at the centre of every AI engine. And data preparation is one of the most critical phases.

It starts with collecting data such as stock prices, fundamental data, indexes futures, currencies, interest rates, macroeconomic indicators, analysts' estimates, and technical indicators from our data providers.

During the data preparation step, our Quants and Data Engineers validate, identify and rectify inconsistencies, and deal with outliers, anomalies, structural changes and any missing data. After the data preparation, the data is clean and ready to be transformed and engineered to create features.

Why is it important?
Data preparation helps us to get the most out of the data. In this step, the quality and the quantity of data influence directly how good our predictive model will be. Without this phase, we can end up spending a lot of time looking at bad AI results. The better the data, the more outstanding the results.

What is the output?
Output: clean datasets, ready to be used to train the AI models.
In our case, for each investment product (Equities, Futures, Commodities), there is a different dataset containing features.

INTRO: This step may be considered the heart of the Axyon IRIS technology stack.

💡If our products were cars, this is the step where we produce the engines. Here, our Machine Learning Team will use our proprietary AutoML engine as a factory to train and optimize forecasting models, based on multiple state-of-the-art AI technologies.

More specifically, the clean dataset generated in the previous step will be used to train supervised machine learning models with the goal to forecast the ranking of target financial assets in terms of expected return.

Training the models means defining an objective function that they will have to maximise (e.g. log-likelihood) and letting them autonomously learn how to do so by trial-and-error from historical data.

💡This is analogous to how you would slowly learn a new language by attempting to translate short random sentences and receiving feedback from a teacher: you would slowly learn what works and what doesn't, and gradually improve at this task.

During the training, each ML model learns how to rank financial assets based on the expected relative performance over a given time horizon after processing millions of historical data points. Besides optimizing individual models, our Platform is able to automatically tune hyperparameters** and perform feature selection, thanks to advanced Evolutionary Computation*** techniques.

**Hyperparameter tuning means adjusting the parameters that the model cannot learn and need to be provided before training.

***Evolutionary Computation (EC) is the technique used in complex optimization on problems that have too many variables for traditional algorithms to consider.

Models are then evaluated in terms of proprietary metrics assessing both their accuracy and, crucially, robustness to several scenarios that may occur. Finally, an ensemble of the best models is automatically created and deployed to our model registry for live usage. We also generate out-of-sample historical predictions for unbiased model evaluation and analysis in this step.

Processing large amounts of data is very computationally demanding. Therefore, our Platform relies on a flexible and scalable workload management system which automatically summons accelerated computation nodes on the cloud (supporting multiple providers) or on connected HPC (high-performance computer) resources.

Why is it important?
This is important for two reasons:

(i) a big part of Axyon's R&D effort is dedicated to the improvement of financial AI modelling, hence this step is where our technology really shines;

(ii) the resulting AI models and their forecasting ability will be one of the key ingredients in the construction of our AI-powered investment strategies in later phases.

Once trained, our AI models will be ready to answer the key question:
"How will asset XYZ perform over the next month compared to other assets?"

How long does this phase take?
In general, training ML models can take from seconds on a laptop to months on a large-scale GPU cluster. It depends on the complexity of the model and dataset.

In our case, it may take from a few days up to about a week, depending on the size of the target investment universe.

Our Machine Learning Team is actively working to improve the ML model training process, with the goal of training more accurate models, faster.

What is the output?
This step produces optimised AI-based forecasting models, which will be deployed as a key component of the end product.

After the model deployment, our proprietary AutoML engine produces AI-predictive signals that allow investors to access AI algorithms insights.

The signals are transformed into rankings based on the assets' predictive performances given a time horizon. Those AI signals are then monitored and analysed to determine whether or not they have helped address the client's business needs.

Why is it important?
This is important to evaluate the overall quality and stability of the signals and to understand the relationship between known investing factors and ML signals.

What is the output?
Output: reports that contain detailed information about the historical signals

The final outputs of the model discovery process are relative performance rankings that can be used to build different investment strategies (according to the client's requirements).

These rankings identify predicted outperformers and underperformers within a given investment universe and time horizon.

The first way to integrate AI into an existing process is directly trading a fully-fledged AI-based strategy.

To build such a strategy, an investor could transform the daily ranking into target portfolio weights that guide the portfolio allocation. As an example, a portfolio manager could choose to create a long-only strategy that every week puts equal weight on each of the top 30 assets of the AI-generated ranking.

A different alternative to integrating AI into an existing process is treating AI signals and rankings as an alternative alpha-generating data source. This approach can be illustrated using predictive rankings for overweight or underweight assets already traded by discretionary asset managers.