At Aisot Technologies, we’re passionate about using data to drive smarter investment decisions. One of the tools we use is the APT model (Arbitrage Pricing Theory), which helps us forecast returns on investments by analyzing a range of factors, jointly. In this post, we’ll break down how the APT model works and explain how we make sure our predictions are transparent and easy to understand.
The APT model is a way to estimate the expected returns of investments by looking at multiple factors that could influence those returns. Think of it as a more general and data-driven version of the well-known Capital Asset Pricing Model (CAPM). The APT model takes into account not just market factor but also:
- Macroeconomic factors: Interest rates, inflation, etc.
- Company-specific factors: Financial statements, earnings reports, etc.
- Industry factors: Performance of specific sectors like technology or healthcare.
- Sentiment-based factors: Public opinion gathered from news or social media, via large-language models.
In simple terms, APT helps us make informed predictions about how different investments will perform, using a broad range of information. We update the model daily to ensure it reflects the latest market conditions and insights.
Uncertainty and combining views - Black-Litterman
Note, that we do treat the outputs of the APT model, as a definite forecasting tool, but rather as a probabilistic view on the instruments future performance. Every view is coupled with its associated uncertainty about the forecasted values. Furthermore, when additional views about the market are available from our clients’ research side, we take them into account by combining them with a Black-Litterman framework. This hybrid Bayesian framework allows us to combine our prior information about the instrument pricing (e.g. CAPM prior) with the views that are coming from machine learning models, giving us more robust forecasts. This allows us to refine our views on individual assets and create optimized portfolios that carefully balance both risk and reward.
Making Predictions Understandable: The Role of Shapley Values
A key part of our process at Aisot is ensuring that the predictions we make are not only accurate but also explainable. Investors need to understand not just what the model predicts, but also why it makes those predictions. To achieve this, we use Shapley values. These are a way to measure the importance of each factor (like market trends or company performance) in the model’s final prediction. Shapley values help us break down the model’s decision-making process and show how each factor contributed to the forecast. This makes our models transparent, so clients can see exactly why we recommend certain investments. For example, if the machine learning model predicts that a stock will perform well, Shapley values allow us to explain how much of that prediction was driven by recent news sentiment, market trends, or industry performance. Furthermore, Shapley values provide local explanations, meaning they assess feature importance for each specific prediction instance. This is crucial when you want to understand why the model made a particular prediction for a specific observation, something that global coefficients cannot explain. Additionally, they take into account the contributions of features in combination with other features, which makes them robust to factor correlations and better at attributing importance when features interact. This helps build trust and clarity around the decision-making process.
AI Insights Platform view: overview of all factors influencing the “Sentinel U.S. Equity” portfolio
AI Insight Platform view: unveiling the influence of different factors on the weights of individual stocks in the portfolio
How APT Works in Aisot's Models
Our APT models use a wide variety of data to make predictions, and we tailor them depending on the type of assets we’re analyzing. Here's a simplified look at what factors we use:
Equity Model:
- Market and industry factors: How specific sectors like technology or healthcare are performing.
- Statistical factors: Advanced methods like PCA (principal component analysis) to analyze patterns.
- Time-series features: Trends over time, like returns and volatility.
- Sentiment-based factors: Insights from news and social media.
Crypto Model:
- Market factors: Both traditional market trends and crypto-specific factors.
- Statistical factors: Similar to equities, using advanced methods to analyze data patterns.
- Time-series features: As with equities, looking at trends like momentum and volatility.
- Crypto-specific features: Blockchain data, such as hash rates or transaction volumes, and derivative market stats like futures or staking analytics.
Tailored Features for Clients
In addition to these core factors, we also offer premium features tailored to specific client needs. These can include things like:
- Ratings and analyst scores
- ESG scores (Environmental, Social, Governance)
- Macroeconomic indicators (interest rates, inflation)
- Fundamental data (company balance sheets, earnings reports)
- Crypto-specific analytics (staking ratios, blockchain statistics)
Conclusion: Delivering Transparent and Data-Driven Insights
At Aisot, we combine the power of advanced APT models with explainability techniques like Shapley values to provide clients with clear, actionable insights. Our approach ensures that predictions are not only accurate but also transparent, so clients can understand the reasoning behind every recommendation.
If you’re interested in learning more about how we use data to drive investment decisions, check out some of the research papers from our team, covering everything from alternative data to sentiment analysis and blockchain insights.
