Analysts utilise financial analysis to dissect and assess a company's financial health. Using financial ratios, they scrutinise current data alongside historical trends to unlock key insights like the company's profitability, liquidity and solvency.
As Santa Clara University notes, the analyst is the bridge between pages of raw financial data and the strategic decisions that drive a business forward. Completing a full financial valuation is likely to take three to five weeks. However, in the last few years, companies have begun exploring machine learning to save time when conducting a financial analysis.
After all, a small yet significant part of financial analysis revolves around time-consuming administrative tasks, such as:
Financial analysis can also involve repetitive tasks like data extraction and classification. Machine learning can automate many financial analysis processes because of its ability to learn from experience and mimic some aspects of human cognition. More recently large language models have also been shown to encode knowledge of finance at a very advanced level. Ultimately, automation reduces the administrative burden on analysts, freeing them to focus on more strategic tasks worthy of their skills.
Inputting financial statements into a generic machine learning tool won’t likely yield an accurate output – a contributing factor to why ChatGPT generates errors. The first step in refining machine learning models for financial analysis is to train them on vast sets of high-quality financial data.To learn more about the technical details of training machine learning models, check out our article ‘How Do LLMs Read PDFs?’
While Optical Character Recognition (OCR) can convert financial statements into digital files, its limited capabilities become clear when extracting from diverse statement formats. Having been trained on certain templates, any deviations result in compromised data quality. However, combining OCR with computer vision, a type of AI, empowers it to extract structured data from specific sections within the statements.
If an error occurs, it’s important to isolate and identify it quickly. Machine learning algorithms can check the internal consistency of the extracted financial data before leveraging it to compute ratios. Hence, the tool should flag the error for manual review if something doesn’t add up. Once the data is validated, the machine learning algorithms will apply this insight to avoid future errors – like a human learning from their mistakes.
For calculating financial ratios, the machine learning model’s analysis must be readily understandable to humans. Like a human, the tool should be able to justify its decisions. For example, it should be clear that the tool has placed ‘inventory’ and ‘accounts receivable’ in ‘Current Assets’ in line with accepted accounting practice.
In the past, any attempt to leverage machine learning to analyse financial statements has generally failed. The reason? Machine learning algorithms struggle to extract the relevant information from financial statements – the first step of analysis.
Financial statements are difficult to extract from for several reasons. Firstly, unlike bank statements and invoices, they lack a predefined page structure. Irina Staneva, formerly an auditor at PwC, explains this further:
‘Financial statements are often part of the annual reporting packages for businesses. Extracting the relevant financial information from reports can be tricky because the format and representation of data vary by industry. Company-specific designs can also impose additional challenges for both analysts and algorithms.’
Also, it is difficult for machine learning algorithms to make the right decision consistently, for example, when classifying tangible and intangible assets on balance sheets. Moreover, since financial statements are often part of larger documents, machine learning technology must have document recognition capabilities.
The advent of large language models (LLMs) demonstrated that AI and machine learning can make versatile and context-dependent decisions in the same way as humans. The problem is that LLMs like ChatGPT often generate errors and hallucinations. Our clients reported that when they used ChatGPT to capture and review data from documents, ChatGPT’s results would contain at least one hallucination per page. ChatGPT’s lack of specialist training means it doesn’t have validation capabilities suitable for documents as complex as financial statements.
In contrast, IDP tools excel at extracting data, but they lack the sophistication needed for in-depth financial analysis. Financial statements are filled with subtle nuances that require a deeper grasp of financial concepts, similar to the capabilities of Large Language Models (LLMs).
Combining the strengths of LLMs and IDP promises an effective machine learning solution. We’ve recently launched Financial Statements AI – a tool for extracting and calculating core ratios from financial statements. Creating Financial Statements AI was a mammoth machine learning task involving excessive time and expertise.
We also run the London Machine Learning Meetup – Europe’s largest machine learning and AI community. Even with academic insights gleaned from our community and the industrial expertise of our engineers, it still took months to build and debug this product.
Our project demonstrates that financial analysis requires a sensitive approach, prioritising accuracy and convenience. Building both into a solution requires extensive dedication and resources when training machine learning algorithms to read financial statements like humans. Companies looking to deploy machine learning to assist financial analysis may do well to bear this in mind.
Our article on Venturebeat explored the emerging field of research dedicated to enhancing business decision-making with large language models (LLMs). At the moment, advanced AI and ML technology can extract data and compute core metrics. What it can’t do, however, is use these insights to make reliable recommendations. Analysts remain the necessary link between financial data and business strategy.
Of course, AI and ML capabilities are improving at an almost unbelievable speed. One application we may see become more dominant is financial forecasting – converting financial analysis into a holistic overview of a company’s past, present and projected future. Accurately aggregating and calculating financial forecasting data can take weeks, yet machine learning has proven to be exceptional at compressing lengthy tasks into minutes.
Advances in machine learning offer an increasingly sophisticated approach to financial analysis. Well-trained ML models can provide insights into financial statements faster (and sometimes more accurately) than human analysts. Behind the scenes, however, machine learning algorithms must be well-trained and constantly monitored to analyse financial data accurately.
To learn more about Financial Statements AI or try it out, please visit our page. Alternatively, you can email our team at hello@evolution.ai.
