Generating Comparative Explanations of Financial Time Series
Generating Comparative Explanations of Financial Time Series
At Politecnico Di Torino I spent one year exploring financial time series summarization techniques under Luca Cagliero in the Computer Science Department, in particular, developing methods to automatically generate explanations of stock price movements using advanced natural language processing and machine learning techniques.
My project involved programming, generating, and evaluating variations of comparative explanations for financial time series data using a novel approach that combines time series embedding representations with natural language generation techniques.
During the course of my research, I developed a system that can generate five different types of comparative summaries:
- Virtuous stocks: Compares a single stock to the most virtuous stocks based on a fundamental indicator.
- Sectors: Compares market sectors instead of individual stocks.
- Years to stock: Indicates how many years one stock has been similar to another.
- Year to years: Compares a single year of one stock with all years of another.
- Virtuous multivariate: Specifies a percentage of reference stocks with a given level of similarity, considering multiple indicators.
The system uses a three-step process to generate these summaries:
- Financial data encoding
- Quantifier/summarizer evaluation
- Protoform generation
I evaluated the system using both intrinsic and extrinsic methods, demonstrating its effectiveness in generating accurate and useful summaries for financial analysis.
After completing the project, I wrote my thesis titled “Generating Comparative Explanations of Financial Time Series” and presented my findings at ADBIS 2022.