Daily AI Trading Paper

Financial sentiment analysis sounds straightforward at first: read a news article, decide whether it is positive, negative, or neutral, and turn that judgment into a trading signal. But in practice, this is one of the hard problems in algorithmic trading. Financial language is subtle, company-specific, and full of context that does not live inside a single article. A headline about weak guidance can mean very different things depending on a firm’s prior filings, business model, competitive position, or recent strategic changes. And in trading, a model that is slightly more accurate is not enough if you cannot explain why it made a call, especially when that call is going to drive a long or short position. This paper tackles exactly that gap. The authors argue that modern finance-tuned large language models have improved sentiment extraction, but they still mostly treat each article in isolation and remain black boxes. That creates two problems. First, the model may miss broader company-level context that matters for interpretation. Second, if a portfolio manager wants to review the signal, there is little evidence beyond the model’s output label. The paper’s answer is FinLLM-KG, a framework that combines a causal LLM with a financial knowledge graph built from SEC 10-K filings. The key idea is simple but powerful: do not retrain the language model, just give it better context at inference time, and make that context explicit and inspectable. To understand why this matters, it helps to think about the landscape. Financial sentiment analysis has gone through several stages. Earlier systems relied on lexicons and shallow classifiers. More recently, finance-specific LLMs have been adapted through supervised fine-tuning, preference optimization, and reinforcement learning. Those systems can be strong, but they still usually read one article at a time. Meanwhile, knowledge graphs have become a way to represent financial facts in a structured, time-stamped form. In finance, that is especially valuable because SEC 10-K filings are standardized, public, dated, and attributable. They can capture things like business relationships, risks, strategy, and financial developments in a way that is easy to inspect. The paper’s central claim is that LLMs and knowledge graphs are complementary: the LLM brings language understanding, while the graph brings explicit company memory and provenance. FinLLM-KG uses a published knowledge graph called FinReflectKG, which contains schema-constrained triples extracted from 10-K filings for more than 700 companies across fiscal years 2014 through 2024. Each fact is stored as a triple tagged with a ticker and fiscal year, and the framework only uses the latest 10-K filed strictly before the article date. That temporal gate is important. It prevents leakage from future filings and keeps the context realistic for trading. If no eligible filing exists, the model simply does not get graph context. There are two ways the company knowledge is presented to the LLM. The first is a natural-language company profile. The graph triples are verbalized into short prose under headings like Business, Financials, Momentum, Strategy, and Key Risks. These profiles are built cumulatively across years, so the context can reflect multi-year trends rather than a single filing snapshot. The second is a relevance-ranked set of factual triples. In that version, each triple is embedded, the article is embedded too, and the top matching facts are retrieved by similarity. The default is a small set of four triples. This is a nice design choice because it gives two different flavors of context: one more readable and narrative, the other more compact and fact-oriented. Crucially, the framework is model-agnostic. The language model weights are frozen, so there is no retraining. The article is scored in one forward pass under three conditions: article only, article plus company profile, or article plus retrieved triples. The model is instructed to use the background as guidance, but to let the current article take precedence if there is a conflict. Because the output is a causal LLM, the paper converts the discrete sentiment label into a continuous score using a logit-to-score method. That continuous score is what becomes the trading signal. The experimental setup is fairly realistic for a sentiment-driven trading study. The authors collect about 200,000 financial news articles from February 2015 to June 2021, along with daily stock returns for the 100 companies in the S&P 100. Since news articles often mention multiple firms, they use named entity recognition to identify the target company and keep only articles where the target company’s confidence exceeds 98 percent. That reduces the corpus to about 160,000 articles. They then compare FinLLM-KG against FinDPO, a finance-specific causal LLM that the paper treats as the state-of-the-art baseline for portfolio construction. Because both methods use the same sentiment scoring procedure, the comparison isolates the value of the added company-level knowledge. From there, they build daily long-short portfolios. On each trading day, companies are ranked by sentiment strength, and the signal is used to determine long and short positions. The universe is dynamic and point-in-time, meaning only companies with available news and market data on that day are included. That matters because it avoids a common backtest mistake: using a static list of names that would not have been investable at the time. The paper’s main result is that FinLLM-KG consistently improves portfolio performance over the baseline. The abstract says simulations show better performance, and the body emphasizes that the framework improves both predictive and trading performance. What should we take from that? Not that knowledge graphs magically create alpha, but that adding temporally grounded company context helps the sentiment model make better distinctions, and those better distinctions carry through into the long-short portfolio. In other words, the improvement is not just a classification gain in the abstract; it shows up in a trading setting where the signal has to survive ranking, portfolio construction, and market noise. The explainability angle is also important. FinLLM-KG is designed to attribute each article-level sentiment score back to individual sentences and to the retrieved knowledge-graph facts. That gives a reviewer a way to inspect which parts of the article and which external facts supported the model’s view. In finance, that is more than a nice-to-have. It supports human-in-the-loop review, helps with model governance, and makes it easier to detect when the model may be leaning on stale or misleading context. There are a few limitations to keep in mind. First, the framework depends on the quality and coverage of the underlying 10-K knowledge graph. If the filings omit something important, or if the extracted triples are noisy, the context can only be as good as the graph. Second, the system still relies on sentiment as the main signal. That works well for event-driven strategies, but sentiment is only one lens on market behavior, and not every price move is sentiment-driven. Third, the paper evaluates on a specific news universe and an S&P 100 trading setup, so the results may not transfer unchanged to smaller-cap names, faster intraday horizons, or non-U.S. markets. Finally, although the framework is explainable relative to a black-box LLM, it is not fully causal in the economic sense; it improves interpretability of the model’s evidence, not necessarily the economic mechanism behind returns. Even with those caveats, the contribution is quite practical. The paper shows a low-friction way to make finance LLMs more context-aware without expensive retraining: retrieve temporally valid company facts, inject them at inference time, and keep the signal auditable. For trading teams, that combination of better context, lower compute, and stronger transparency is exactly the kind of design that can move a sentiment model from a research demo toward something more deployable.