Paper: arXiv 2309.03079
Authors: Udit Gupta
Abstract
Annual Reports of publicly listed companies contain vital information about their financial health which can help assess the potential impact on Stock price of the firm. These reports are comprehensive in nature, going up to, and sometimes exceeding, 100 pages. Analysing these reports is cumbersome even for a single firm, let alone the whole universe of firms that exist. Over the years, financial experts have become proficient in extracting valuable information from these documents relatively quickly. However, this requires years of practice and experience. This paper aims to simplify the process of assessing Annual Reports of all the firms by leveraging the capabilities of Large Language Models (LLMs). The insights generated by the LLM are compiled in a Quant styled dataset and augmented by historical stock price data. A Machine Learning model is then trained with LLM outputs as features. The walkforward test results show promising outperformance wrt S&P500 returns. This paper intends to provide a framework for future work in this direction. To facilitate this, the code has been released as open source.
Complexity vs Empirical Score
- Math Complexity: 2.5/10
- Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; see penalty flags)
- Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.5)
Why this score: The paper relies on standard machine learning and statistical measures without advanced mathematical derivations, but features a detailed backtest with walk-forward analysis, transaction cost considerations, and open-source code, making it implementation-heavy.
Research Flowchart
flowchart TD A["Research Goal:<br>Automate Annual Report Analysis<br>for Stock Investment"] --> B["Data Collection:<br>S&P500 Annual Reports &<br>Historical Stock Prices"] B --> C["LLM Analysis:<br>Extract Insights & Sentiment<br>from 100+ Page Documents"] C --> D["Dataset Construction:<br>Combine LLM Outputs with<br>Quantitative Financial Data"] D --> E["Machine Learning:<br>Train Predictive Model<br>on Augmented Dataset"] E --> F["Validation:<br>Walkforward Testing<br>vs S&P500 Benchmark"] F --> G["Outcome:<br>Significant Outperformance<br>with Open-Source Framework"] style A fill:#e1f5fe style G fill:#e8f5e8 style F fill:#fff3e0
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes extracting features from annual 10‑K reports using a large language model and training a machine‑learning model on these LLM outputs combined with historical price data to select stocks.
Source sentence
The insights generated by the LLM are compiled in a Quant styled dataset and augmented by historical stock price data. A Machine Learning model is then trained with LLM outputs as features.
Assumes 3
- The study is limited to the top 1500 US companies by market cap that file 10‑K reports. market
Source sentence
The stocks considered for this paper are from the US stock universe. Specifically, the top 1500 companies by market cap in the US which are part of the S&P 500 largecap index, S&P 400 midcap index and S&P 600 smallcap index.
- Annual‑report‑based signals are assumed to have a longer investment horizon than short‑term news‑driven signals. execution
Source sentence
As these reports are published annually, the signals generated from these are of longer duration and stock selection using these signals…
- Training data covers 2002‑2017 and testing data covers 2018‑2023. data
Source sentence
In this exercise, data from 2002 to 2017 is considered for training the ML model, and data from 2018 to 2023 is considered for testing.
Data
- Historical 10‑K filings fundamentals public
Source sentence
To construct the dataset, access to historical 10-K filings of the top 1500 companies, ranked by market capitalization, is required.
- Historical stock price data market-prices unknown
Source sentence
The insights generated by the LLM are compiled in a Quant styled dataset and augmented by historical stock price data.
Costs
Not modeled
Finds 3
- The ML model built on GPT‑3.5‑derived features yields higher returns than the S&P 500 over the test period. performance
Source sentence
The analysis of predictions made on the test set reveals that the Machine Learning model constructed using GPT 3.5 features outperforms in selecting stocks, resulting in returns that surpass those of the S&P 500 Index.
- Both the 12‑month‑target and max‑return models outperform the S&P 500 from 2018 to 2023, with the max‑return model showing substantially higher gains. performance
Source sentence
Figure 3 depicts the investment performance of $1 from the start of 2018 to the start of 2023 for two differently constructed models. While both models outperform the S&P 500 returns during this period...
- The max‑returns strategy can quadruple a $1 investment over a five‑year horizon. performance
Source sentence
Figure 4b shows the best case scenario where the Max Returns strategy is able to quadruple the $1 investment in 5 y…
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 10 related papers). Coverage is partial, so this is neither replication nor refutation.
8 papers in the archive : Language of Persuasion and Misrepresentation in Business …
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.