How AI Delivers Investment Information

Toss Securities is developing financial investment information services, including AI signals, real-time issues, earnings calls, and an AI stock market calendar, leveraging ML and LLMs. Unlike news summaries or disclosure rewrites, investment information must simultaneously satisfy timeliness, accuracy, verifiability, and anomaly detection.

In financial markets, the interpretation of the same information varies depending on the timing and market conditions. The mere appearance of a company name in an article does not constitute a basis for stock price movements, and all model outputs must be traceable and reproducible. LLMs and autonomous agents can generate fluent but incorrect answers when information is missing or erroneous, and their execution paths vary with each run.

Accordingly, Toss Securities defined three gates that AI must pass before providing an answer.

• Choosing what to say: Selecting the evidence to pass to the LLM through source data classification and retrieval, deduplication, reranking and filtering, and context alignment.
• Controlling how to generate: Combining retrieval, classification, verification, and evaluation into a predefined task graph, focusing the LLM on summarization and evidence-based explanations.
• Creating an evaluable structure: Establishing a structure that allows for observing and improving results during operation, rather than just creating prompts once.

In the first gate, metadata is assigned using a BERT-based classification model from the moment news, disclosures, and financial data arrive. It classifies relevance to Korean stocks, whether analyst analysis is present, technology trends, and related companies, while also pre-generating embeddings for vector search.

Subsequently, candidates are broadly collected using a hybrid retriever, duplicate information is removed via semantic-based clustering, and reranking is performed based on direct relevance to corporate stock price movements. Explanation patterns and failure reason taxonomies are applied together to distinguish between earnings, guidance, corporate actions, promotional information, and lack of evidence.

The final evidence is arranged for the LLM to check in the following order: what happened, how the event connects to the company, whether the price direction and evidence polarity align, and whether the evidence is sufficient and up-to-date. This constitutes a Reasoning-Friendly Context.

In the generation stage, autonomous agents and procedural orchestration are distinguished based on service requirements. For tasks with open problem definitions, such as discovering investment ideas, agent exploration and planning are useful. However, for tasks with clear response requirements, such as explaining stock price fluctuations, a task graph providing a reproducible execution path is more suitable.

Running a long ReAct loop increases tool calls, token usage, latency, and traces that need to be managed. Therefore, in products requiring a defined decision path, the input and output schemas for each step—from candidate retrieval to relevance judgment, deduplication, evidence construction, and final response generation—are clearly defined to enable debugging, evaluation, and cost, speed, and fallback design.