Why do firms hoard cash?

Using AI to rethink financial constraints

By Rachel Cho, Danny McGowan, Christoph Görtz and Max Schroeder

Like squirrels storing nuts for winter, some companies save part of their cash flow for tougher times. Economists often point to financial constraints to explain this phenomenon: firms anticipating difficulties raising money from banks or investors will set aside some of their earnings for future use. These constraints can shape business decisions on investment, hiring, and cash management, with knock-on effects for the wider economy.

For this reason, understanding how financial constraints arise, and how firms deal with them is an important question. However, financial constraints are tricky to measure. Traditional methods rely on accounting data published long after decisions are made. More recent approaches have used company filings, searching for words linked to financial pressure. Yet these methods often miss the subtle ways firms talk about money. They may also fail to capture whether a problem is happening now or is expected in the future.

Our new ESCoE discussion paper develops a new way to tackle this challenge, using artificial intelligence to read corporate disclosures more carefully and understand the nuances of corporate communications. By training a language model (BERT) to understand how firms describe their financing conditions, we can distinguish between those that are currently constrained and those that anticipate constraints in the future. This timing dimension turns out to be crucial for understanding when firms choose to hoard cash.

How can we teach AI to read corporate reports?

The paper focuses on the Management Discussion and Analysis (MD&A) section of 10-K filings, where US-listed companies are required to explain their financial health, risks, and liquidity plans. It’s here that managers often discuss whether they face difficulties raising funds.

To help us analyse these disclosures, we turn to BERT, a state-of-the-art AI language model. BERT is well suited to this task because it processes whole sentences, rather than scanning for individual keywords. This means it can pick up on more context and nuances. For example, this helps it tell the difference between a firm saying it expects financing difficulties next year and one saying it is currently unable to borrow. By fine-tuning BERT on examples from 10-K filings, we train it to recognise and classify statements about financial constraints with far greater accuracy than traditional methods.

To train our model, we followed an active learning approach (see figure below). Starting from a small sample of human-labelled reports, we fine-tuned BERT to identify text that signals financial constraints. The model then suggested new passages for human reviewers to check – especially those it found difficult to classify. Over several rounds, the system improved, building a training dataset of nearly 24,000 labelled text segments. By the end, our model could classify constraint-related disclosures with a 94% accuracy.

We then extended the model to classify constraints along two further dimensions:

• Severity: Mild, moderate, or severe.
• Timing: Whether the constraint is current, future, or both.

This allows us to move beyond a simple “constrained/unconstrained” view and instead considers the severity of the constraints, and when they are binding.

How do anticipated vs. current financial constraints shape firms?

One striking pattern is that most discussions of constraints are forward-looking. Mild and moderate constraints are almost always described as anticipated rather than current. But among firms facing severe problems, nearly half say they are already experiencing constraints.

This matters because theory suggests that firms behave differently depending on whether they are worried about the future or already struggling today. When expecting trouble ahead, firms may start saving cash now as insurance. But if already in trouble, they may have little room left to save.

We tested this hypothesis by revisiting a classic idea in corporate finance: the “cash flow sensitivity of cash”.¹ The idea is that firms anticipating constraints should save more of each dollar of cash flow. But past studies have struggled to pin down this effect consistently.

Using our AI financial constraints indicators in conjunction with firms’ balance sheet data, we conduct several econometric tests to determine the importance of constraint timing on saving out of cashflow. Our results show:

• Firms anticipating future constraints do indeed hoard cash, saving a portion of their internal cash flow.
• Currently constrained firms, by contrast, do not. Their financial position is already too tight to allow precautionary saving.
• Unconstrained firms can still access capital markets and have no reason to stockpile cash.

This new evidence lines up with economic theory and highlights the importance of not only distinguishing whether firms are constrained, but when.

Why does AI-driven economic measurement matter for policy, firms, and finance research?

Our approach shows how AI can enhance economic measurement in ways that traditional data sources cannot. By having BERT read firms’ own words, we gain a nuanced picture of financial frictions. This has at least three broader implications:

1. Policy monitoring – Modern AI gives regulators and policymakers the tools to track emerging financial pressures earlier and more accurately, spotting when firms expect constraints to tighten before this shows up in hard data.
2. Corporate behaviour – Understanding the timing of constraints helps explain why some firms accumulate large cash buffers while others run them down.
3. Conceptual clarity – By resolving past ambiguities around the cash flow sensitivity of cash, we provide a clearer foundation for future work on corporate finance and macroeconomic stability.

Conclusion

Financial constraints have long been a puzzle for economists: they are hard to observe, yet are critical for shaping corporate behaviour. By applying AI to company filings, we show that it is possible not only to identify which firms are constrained, but also to separate those that are worried about the future from those already struggling today.

The distinction turns out to be vital. Firms save cash as a precaution only when they anticipate trouble ahead – not once they are already constrained. This insight brings theory and evidence into alignment and opens new possibilities for monitoring financial health in real time.

ESCoE blogs are published to further debate.  Any views expressed are solely those of the author(s) and so cannot be taken to represent those of ESCoE, its partner institutions or the Office for National Statistics.

¹ Almeida, H., Campello, M., & Weisbach, M. S. (2004). The cash flow sensitivity of cash. The Journal of Finance, 59(4), 1777–1804. https://doi.org/10.1111/j.1540-6261.2004.00679.x