The Science Behind Sphere

Introducing Sphere

MDOTM is one of the leading Fintech companies globally for the application of Artificial Intelligence (AI) techniques to financial markets. Since its foundation in 2015, the company has been a pioneer in the development of investment strategies that are fully based on AI, offering institutional investors a suite of bespoke solutions that support them in their investment process.

The increasing complexity of financial markets has turned technology into the key to extracting information hidden in the huge volume of noisy background data. For this reason, MDOTM has developed a proprietary AI methodology called ALICE® (“Adaptive Learning In Complex Environments”), which exploits changes in market inertia and risk premia to enable robust and reliable investment decisions. sophisticated deep learning techniques allow ALICE to daily analyze millions of market data points and learn how to build efficient and diversified portfolios.

To respond to the growing need of making more-informed investment decisions, MDOTM has developed an AI-driven platform called Sphere®  that allows institutional investors to directly implement and interact with its proprietary technology, giving them the possibility to monitor markets, create, analyze, and manage portfolios and build investment strategies guided by AI-driven insights. The goal is to simplify and empower the work of investment professionals, helping them to efficiently interpret and collect the correct insights from the huge amount of data in the market, ultimately enhancing investment decisions. This helps them gain a competitive edge while staying at the forefront of innovation.

In this document we will be presenting how the technology behind Sphere works, going through MDOTM’s approach to research and the model guiding its decision-making process. The aim is to shed light on the methodology used and to show why this Platform can effectively assist and add value to client’s investment decisions.

Start integrating AI into your investment process with the most advanced no-code solution.

Fill out the form to book a demo of Sphere®, and learn how it can bring value to your investment decisions.

Book A Custom Demo

By submitting you are agreeing to the Terms of Use and Privacy Policy
Thank you! Your submission has been received.
Oops! Something went wrong while submitting the form.

Our Approach to Research

MDOTM’s approach to investments is deeply rooted in the scientific method, relying on extensive and rigorous research carried out by both its internal R&D team (50+ people) and in partnership with professors, PhDs and students via MDOTM Lab – MDOTM’s research center.

When conducting research, the first step consists of a deep review of the existing academic literature (or, alternatively, an analysis started from scratch) of a possible market phenomenon that can be exploited to create value. From the “raw” effect described in the literature or identified in its preliminary analysis, MDOTM refines its models with incremental improvements to the extent that this effect is found to add value to the model’s risk-adjusted performance or statistical significance of results.

To conduct the research process, MDOTM uses different advanced machine learning techniques, such as:

  • Random Forests classification algorithms, as part of the selection process of the features/variables that are relevant to the models;
  • Algorithms to identify fitness functions that favor stable solutions through a Grid Search tool - to make hyper-parameter optimization - and the Multivariate Surface Analysis - which is an advanced data exploration module - to carry out hypothesis testing
  • Neural Networks - to perform parameters clustering analysis;
  • A derivation of Hidden Markov Models (HMM) which is an unsupervised algorithm to make forward-looking probabilities forecasts - to provide an unbiased estimate of the expected market regime.

All the research conducted follows MDOTM’s rigorous belief that to get robust and reliable results, it is not “the more data, the better results”, but rather the quality of the data that plays a crucial role in establishing the quality of the output. This also applies when evaluating the optimal time horizon for which data can be considered relevant.

The input data from which our AI learns can be divided into three main categories:

  • Historical market data, which is unbiased, indisputable, and reliable. It is also consistent and has been available for a long time. Using the historical market returns, MDOTM performs different types of analysis, such as the cross-sectional study of the returns, the analysis of return distributions, and the study of asset correlations - using var-covar matrices and cointegrations.
  • Macro data such as data related to GDP, inflation, and interest rates, that allows one to get a deeper knowledge of the historical market dynamics and evolution.
  • Fundamental Data, studied on an aggregate level (not looking at single companies), to get insight into the structure of the market and investment factors.

Simultaneously, MDOTM also uses descriptive data - such as index composition, sectors, industries, and market cap distribution - to get a better understanding of the historical evolution of markets’ structure.

The data fed into our models are characterized by high predictive power and high predictive horizon, consistently with the fact that MDOTM develops long-only strategies. Therefore, data like news, market sentiment or alternative data is not used as input to MDOTM’s models due to its low predictive power for long-term horizon strategies, also considering that this data can easily be subject to biases of interpretation, and consequently lose its significance.

Feeding its model with this quality data, MDOTM can provide a view of the market’s condition and advise on optimal positioning. The results are obtained via a unique and unbiased approach, resulting in complementary insights that can easily be added to its clients’ existing investment process.  

Figure 1: Above, we can see Sphere’s suggested positioning for the main asset classes compared to the previous one (last 30 days). The Outlook sections provide further insights on the suggested positioning in terms of levels of exposure.

Our Investment Decisions Model

The Human decision process is driven by two key components:

  • Experience: we combine past events in a non-linear way through our imagination;
  • Data Evaluation: we look at data to have the “best” possible understanding of the status quo.

When it comes down to making an investment decision, an experienced portfolio manager (PM) usually follows the same path: to make an educated guess on what could happen in the future, the PM would start by remembering different past market phases, past market dynamics, past black swan events and then would look at today’s context, gathering as much information as possible to combine it in a “non-linear” way.

Artificial intelligence also utilizes Experience and Data Evaluation, but differing from humans in how it does it: AI does it through leveraging data. This is why such an approach pairs up very well with a more traditional/fundamental one. Humans have an edge when information is scarce but AI, on the contrary, has an edge when information is abundant and when it is a matter of connecting and analyzing thousands of data points altogether.

MDOTM leverages its proprietary AI methodology to provide actionable investment insights based on a joint evaluation of the following components:

  • Experience, which is gained through a methodology that we called Non-Chronological Learning: a clustering analysis of past market dynamics and events.
  • Data Evaluation, which happens through a methodology that performs Regime Analysis: a derivation of HMM (Hidden Markov Model) that allows us to assess in which type of risk framework we are in – see below.

1. Non-Chronological Learning

The process to build efficient and robust portfolios starts with experience: our AI models need to be trained to learn from past data. This learning process goes deeper than the mere memorization of past dynamics and occurrences and it is rather based on learning the market properties of past events – i.e. generalization. This means that the model will learn from what has happened in the past, gaining the ability to extract the essence of a concept based on the analysis of similarities from many previous scenarios and events. By doing so, it is then able to adapt properly to new, previously unseen data, and make accurate predictions.

Sphere’s learning process can be explained by means of an analogy: the way humans learn how to recognise an apple tree. Starting from looking at a group of different trees, our first instinct will be to categorise their main characteristics, creating unwritten rules that make it easier to identify them. To do so, we will look at its main features, such as the shape of the leaves, the size of the crown, the shape and colour of its flowers and fruits, among many other possible details. Through this process, we will naturally create groups in order to cluster the different types of leaves, crowns and fruits that we have observed and which we have learned to recognise as parts of a tree.

When faced with a new tree, humans will use the learning process described above to try to identify and classify it, based on the knowledge previously acquired. If the characteristics gathered initially were significant and representative of the reality, the clusters created should include all the necessary details to allow for a new and never previously seen tree to be recognised as such.

DiagramDescription automatically generated
Figure 2: Simplified Clustering Process undertaken to create a probabilistic forecast on expected returns and var-covar matrices.

Our AI model training is made by means of neural networks that observe past market scenarios and cluster them into analogous information groups. When faced with a new scenario, the model recalls the learning process and generates a forecast of the correlation and volatility, whilst evaluating the riskiness of the asset classes, to create portfolios that take into account the future evolutions of the Var-Covar matrix and the tail risk of the relevant assets. Then, combining experience with the “data evaluation” part of the process, the model analyses the current market scenario - identified by HMM regime analysis, as detailed below - and interprets this data in the light of the clusters previously created through “experience”. By doing so, it is possible to build a portfolio that discounts any possible changes in the asset correlation based on the previous market experience. In this process, the neural network gets trained to obtain higher portfolio efficiency, by maximising returns and minimising volatility at every rebalance. As a result, the allocation is not a simple weight optimization reflecting the current scenario, but a portfolio in which single instrument weights are determined by systematically discounting possible changes in the variance-covariance matrices and in market conditions.

Figure 3: Client’s Portfolio details. Here it is possible to consult the portfolio metrics, composition, and its alignment to MDOTM’s views of optimal positioning.

Figure 4: Scenario Analysis module is also available, showing how the portfolio would have performed during specific timesin history.

2. Regime Analysis

Everyday financial markets are in a defined regime characterized by the existing dynamics, conditions, and behavior of different components and variables. In today’s context of growing complexity, it is crucial to be ”regime-aware” when making investment decisions as it allows better positioning of the portfolio’s risk.

MDOTM’s Market Regime Analysis is done by using a derivation of Hidden Markov Models (HMM), an unsupervised machine learning technique generally used to estimate forward-looking probabilities and to understand the evolution of complex systems that cannot be directly observed. The HMM bottom-up approach to market segmentation allows identifying market regimes in terms of risk environment and data coherence, adding value compared to mere return/growth measurements.

Having identified the different regimes, our model looks at the current market structure and compares it to the historical market regimes identified, thus assessing to which specific market regime the current scenario belongs to and how long it is supposed to last. Simultaneously, the technology is also able to look forward and to determine the weighted probability of occurrence of each market regime in the near future, as well as the related transition probabilities considering how long we have entered/been in the current regime. Depending on the market phase identified by HMM, positioning decisions will differ. This is due to the fact that each regime and its underlying characteristics will originate different signals regarding which parameters are more or less significant, depending on the inner coherence/incoherence of the data that is detected during the scenario identification phase.

Figure 5: Regime Analysis Chart in Sphere’sOutlook Analysis. Outlook Analysis of each asset class, showing its behaviour during historicalperiods. The analysis examines the return of a standard portfolio (50% equity –50% fixed income composition) over the last 15 years and highlights periodscomparable with the actual market regime detected by the Hidden Markov Model(HMM).

Summing up, combining Non-Chronological learning and Regime Analysis allow to create a probabilistic forecasting of the relevant portfolio components (such as expected returns and var-covar matrices) whilst embedding a risk management layer into the investment methodology itself rather than as containment of discretionary bets.

Sphere leverages the most powerful AI techniques to support its users in every step and stage of their investment process, in a no-code environment, while ensuring robustness and explainability of investment decisions. It empowers human ability to generate new ideas, turn signals into investable portfolios and effectively enhance existing allocations.

Start Integrating AI Into Your Investment Process