Financial KPI extraction using Large Language Models
(2026) In Master's Theses in Mathematical Sciences FMAM05 20252Mathematics (Faculty of Engineering)
- Abstract
- Extracting financial Key Performance Indicators from annual reports is one of many
manual tasks in the banking sector. This can be automated with Machine Learning techniques such as Question Answering, Named Entity Recognition, Entity
Relation, Text-to-Sql, or Retrieval Augmented Generation. In this thesis, different models performances are evaluated when extracting KPIs from annual reports,
both through a literature review and training custom models. It further investigates whether a custom trained model outperforms current existing state-of-the-art
models. The thesis shows that a custom trained extractive model yields an Exact Match of 53.1 - 59.8%. The result is based on a training dataset created from
annual reports in XHTML... (More) - Extracting financial Key Performance Indicators from annual reports is one of many
manual tasks in the banking sector. This can be automated with Machine Learning techniques such as Question Answering, Named Entity Recognition, Entity
Relation, Text-to-Sql, or Retrieval Augmented Generation. In this thesis, different models performances are evaluated when extracting KPIs from annual reports,
both through a literature review and training custom models. It further investigates whether a custom trained model outperforms current existing state-of-the-art
models. The thesis shows that a custom trained extractive model yields an Exact Match of 53.1 - 59.8%. The result is based on a training dataset created from
annual reports in XHTML format and a test dataset based on the same annual
reports in PDF format. It compares the extractive, custom trained model, to state-of-the-art models, such as ChatGPT and GPT-5-nano. The latter model scores an
Exact Match of 58%. However, regarding the generative models, the results should
be seen as a lower limit, due to minimal effort given towards prompt engineering
and fine-tuning. During training, given the correct context, the custom trained extractive model achieves 95.2% Exact Match on the validation set. This shows high
potential and indicates that an improvement in retrieving the correct context can
significantly increase the model’s end-to-end performance. The thesis also compares
the suitability of different model architectures when extracting financial KPIs from
annual reports. (Less) - Popular Abstract
- You can extract Key Performance Indicators such as revenue and
profit from annual reports using Large Language Models with an
accuracy between 51.0% and 78.5%. Training results show a possibility of achieving as high as 95.2% accuracy.
The experiments and methods behind these results are presented in this thesis.
Every company presents their annual report at the end of their financial
year. It can contain more than 300 pages. Among them are financial Key
Performance Indicators. In order to extract these values, you must first
find the correct table. Then once you have the correct table, the correct
Key Performance Indicator can be extracted. These two problems are what
I tackle using Machine Learning in this thesis. The first... (More) - You can extract Key Performance Indicators such as revenue and
profit from annual reports using Large Language Models with an
accuracy between 51.0% and 78.5%. Training results show a possibility of achieving as high as 95.2% accuracy.
The experiments and methods behind these results are presented in this thesis.
Every company presents their annual report at the end of their financial
year. It can contain more than 300 pages. Among them are financial Key
Performance Indicators. In order to extract these values, you must first
find the correct table. Then once you have the correct table, the correct
Key Performance Indicator can be extracted. These two problems are what
I tackle using Machine Learning in this thesis. The first problem requires
search techniques to find the correct table among the large number of tables. Therefore, three different search techniques were analyzed. The second
problem requires a good model and dataset. I have used existing models and
improved them on the specific task. An important note is that during training, 95% accuracy was achieved. This is a simplified version where the model
is given the correct table, e.g. income statement, balance sheet or cash flow
analysis, depending on where in the annual report the KPI is presented.
The thesis focus on training these models and investigates how to achieve
the highest precision. It also discusses which model architectures that are
most suitable. The conclusion is that the smaller models also called BERT
models, compared to GPT models, perform similarly even with regard to
their smaller size. They are also easier to train and adjust on a specific task.
In order to train these models, a dataset was created from open source
databases. There are vast amounts of similar data that can be downloaded
to train these models. I had to draw a line somewhere. Therefore, the annual reports used in this thesis comes from companies reporting according
to European Single Electronic Format, has Sweden as home member state,
and are written in Swedish.
To conclude, it is possible to extract KPIs and this thesis serves as a good
first step into automating the process. The main benefit would be to alleviate time from manual processes so analysts and employees can focus on more
critical tasks such as customer relation and analyzing the KPIs. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9221154
- author
- Tobrand, Vidar LU
- supervisor
- organization
- course
- FMAM05 20252
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Transformers, LLM, KPI, Information Extraction
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- 2026:E3
- ISSN
- 1404-6342
- 3603-2026
- language
- English
- id
- 9221154
- date added to LUP
- 2026-02-17 09:40:07
- date last changed
- 2026-02-17 09:40:07
@misc{9221154,
abstract = {{Extracting financial Key Performance Indicators from annual reports is one of many
manual tasks in the banking sector. This can be automated with Machine Learning techniques such as Question Answering, Named Entity Recognition, Entity
Relation, Text-to-Sql, or Retrieval Augmented Generation. In this thesis, different models performances are evaluated when extracting KPIs from annual reports,
both through a literature review and training custom models. It further investigates whether a custom trained model outperforms current existing state-of-the-art
models. The thesis shows that a custom trained extractive model yields an Exact Match of 53.1 - 59.8%. The result is based on a training dataset created from
annual reports in XHTML format and a test dataset based on the same annual
reports in PDF format. It compares the extractive, custom trained model, to state-of-the-art models, such as ChatGPT and GPT-5-nano. The latter model scores an
Exact Match of 58%. However, regarding the generative models, the results should
be seen as a lower limit, due to minimal effort given towards prompt engineering
and fine-tuning. During training, given the correct context, the custom trained extractive model achieves 95.2% Exact Match on the validation set. This shows high
potential and indicates that an improvement in retrieving the correct context can
significantly increase the model’s end-to-end performance. The thesis also compares
the suitability of different model architectures when extracting financial KPIs from
annual reports.}},
author = {{Tobrand, Vidar}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Financial KPI extraction using Large Language Models}},
year = {{2026}},
}