ABEJA Researcher Takaito Takano: Corporate Lifecycle Cannot Be Captured on a Single Axis—Text Embeddings Enable Multi-Path Representation

Takaito Takano (Seikei University / ABEJA) presented a co-authored study at the 37th SIG-FIN: building embedding vectors of corporate maturity from text in Japanese securities reports. The starting point was a concrete modelling dilemma—corporate lifecycle cannot be compressed onto a single axis.

The traditional approach, Dickinson (2011), classifies firms into four discrete stages—introduction, growth, maturity, decline—by combining the signs of operating, investing, and financing cash flows. Takano identified two problems: discrete labels cannot express variation within a stage or continuous transitions between stages; relying solely on financial indicators makes estimation unstable. They had previously tried treating maturity as a continuous value (a regression from 1 to 4), but found that firms can move from growth directly to decline without passing through maturity, and a single axis cannot encode this skip path.

The solution is to represent maturity as an embedding vector. Text inputs come from three sections of Japanese securities reports (business description, sustainability initiatives, dividend policy). Teacher labels are defined by cross-referencing financial conditions across four dimensions (market and business environment, financials and cash flow, dividends and shareholder returns, organization and governance); firms meeting all criteria for a given stage receive that label. Training uses Triplet Loss: texts of the same maturity attract in vector space, texts of different maturity repel.