Economics 477 | Jasper Paez | December 15, 2025

A study on the impact of Information Technology spending on modern financial firm profit metrics, with a focus on bank-to-fintech comparisons. 

Robert Solow, an economist from the mid-1900s, stated his disappointment in the American economy during an age of rapid technological advancement. Coined from the above quote, the "Solow Paradox" refers to a period of widespread technological innovation, closely followed by a period of stagnating productivity and profitability (Figure 1). In the late 90's, productivity rebounded and disproved the paradox, however its implications still apply today. 

As a Computer Science and Economics student highly interested in Artificial Intelligence (AI) and technology in general, the discovery of the Solow Paradox was the catalyst for this project. 

I began to wonder, with all the talk of rapid AI investments and a supposed "AI bubble", has there been a return? If not, when will returns occur, if ever? 

Let's dive into the details... 

Not only are banks and Fintechs especially (Figure 4) spending more on IT as a percent of expenses, but they are increasingly bringing IT employees in-house (Figure 3). This is paralleled by an explosion in AI advancement and subsequent investment by banks and Fintechs alike (Figure 2). 

So.. Is IT spending a profit driver or just a cost of survival?

and

Do Fintechs (early AI adopters) actually “do it better”?

Research Process

Data Aggregation/Cleaning→ Methodology → Results → Implications 

To begin the study, it was clear that I would need a dataset that did not exist currently. Therefore, I would need to create the dataset from scratch. As seen in Figure 2, the AI boom began around 2017. Therefore, it was important to select a timeframe that included years prior to AI adoption, during its advancement, and all years after. 

To start, I began by first selecting the time period, a number of banks to act as the control group, and a group of Fintechs to act as the treatment group. 

As a result I chose to study 12 banks and 8 fintech firms from 2015 - 2023. The banks ranged from large firms to smaller traditional banks and included global companies, while the Fintechs also varied in size and strategy. 

Starting from the SEC EDGAR database, which offers access to financial documents for all publicly-owned firms, financial data for all firms was gathered from their 10k forms and aggregated into a panel style dataset. Therefore, the dataset has multiple lines for each firm corresponding to their financial metrics for each year from 2015 -2023. 

To supplement the financial data and strengthen the explanatory power of the models, relevant Macro-Economic control variables were collected from the Federal Reserve Economic Database (FRED). 

Finally, to understand the relationship between technology and profit, I created three primary focus variables. Here is how they were constructed and why they matter:

The Data Cleaning Process: From 10-Ks to Insights

1. Comparing IT Metrics: Banks and Fintechs use entirely different accounting terminology and practices, which was a significant issue to address. 

   • • • My Problem: Traditional banks (like JP Morgan) list "Data Processing" or "Communications" as expenses. Fintechs (like Palantir) often bundle their technology costs into line items like "R&D" or "Cost of Revenue."

       • My Solution: I carefully mapped these categories to the 3 expense variables I created, to ensure Tech Spend represented the same functional investment across both groups. This ensured that a Fintech’s "innovation spend" is comparable to a bank’s "modernization spend."

2. Currency & Inflation Normalization: Because I included global firms like HSBC (UK) and Deutsche Bank (Germany), the raw data arrived in various currencies and "nominal" dollars.

   • • • Currency: All foreign filings were converted to USD using the exchange rate at the end of each respective fiscal year.

       • Inflation: To prevent "nominal growth" from skewing the results, I also adjusted all dollar amounts using Consumer Price Index (CPI) data from FRED. This ensured that 2015 spending is directly comparable to 2023 spending in "real" terms.

3. Filtering Out the "Credit Noise" (The Provision Adjustment): A bank's bottom line is often dictated by loan performance, not technology.

   • • Credit Loss: I identified and isolated Provisions for Credit Loss (cash set aside for defaulted loans). By controlling for this, I ensured that a sudden spike in unpaid loans wouldn't "blame" the IT department for a drop in profit margins.

4. Panel Data Structure: I built an unbalanced panel, which allowed me to include newer Fintechs that haven't been public for the full 9-year window, maximizing the data points available for my treatment group. This was interesting, as the study was able to include companies at different maturity stages, and therefore widely different levels of IT investment. 

   • • Handling Outliers: If a firm was acquired or had a massive structural change (like a merger), I adjusted the data to ensure the "Tech Intensity" ratio remained a consistent reflection of the underlying firm’s strategy rather than an accounting mistake.

Regression Design & Methodology

To move from the raw dataset to economic insights, I constructed a series of six regression models designed to isolate the impact of IT spending from the "noise" of the broader economy. I utilized Panel Data methods in STATA, which allowed me to track firms over time rather than just looking at a single snapshot. The process began with a Pooled OLS baseline to establish the general industry trend, followed by Fixed Effects (FE) models to control for firm-specific characteristics that don't change over time (like corporate culture or brand prestige).

A majority of the models focus on Net Profit Margin (Net Income / Revenue) as the dependent variable, with Tech Intensity and controls as the independent variables. To supplement this, one of the models tests Revenue Per Employee as the dependent variable in order to analyze the impact of IT spending on firm financial efficiency. 

To test the "Solow Paradox", my modeling strategy evolved in three stages:

• • • The Interaction Stage: I used interaction terms (Tech_Intensity X is_Fintech) to determine if digital-first Fintechs were achieving a different ROI than traditional banks.

    • The Time-Impact Stage: Recognizing that IT benefits are rarely immediate, I implemented 1-year lags on Tech Intensity to test for the "J-Curve" effect — checking if last year’s investments finally pay off today. 

    • The Efficiency Stage: Finally, I shifted the lens from "Profit" to "Operational Efficiency" by modeling tech spend against Revenue per Employee. This allowed me to see if technology was actually making workers more productive, even if those gains were being swallowed up by "dual-staffing" costs or rapid growth strategies.

Study Assumptions

While the models provided strong evidence of a J-Curve and "Dual-Staffing" trap, I think it is important to acknowledge certain limitations:

Accounting Variance: Despite my "Apples-to-Apples" fix for converting currency, accounting standards for IT reporting are not uniform. Some firms may capitalize software costs (putting them on the balance sheet), while others expense them (putting them on the income statement), which can slightly skew "Tech Intensity" ratios.

The 1-Year Lag Limitation: My study utilized a 1-year lag to test for the J-Curve. In reality, large-scale digital transformations in global banks can take 3–5 years minimum to realize full gains. A longer time interval might reveal a more pronounced recovery in the J-Curve.

Survivorship Bias: The dataset consists of publicly traded firms (SEC EDGAR). This inherently excludes smaller fintech startups that may have failed during the study period, potentially making the Fintech sector look more stable than it is.

Omitted Variable Bias: While I controlled for GDP and Interest Rates, other factors—such as specific regional regulations or internal management shifts—could also influence profit margins independently of technology spending.

The Verdict: IT Spending is A "Survival Tax", Not a Profit Engine

The results of this study suggest that for the modern financial firm, Information Technology spending is currently a strategic necessity rather than a competitive advantage for firms. In economic terms, we are seeing IT spend transition from "Alpha" (a way to beat the market) to "Beta" (the cost of simply being in the market). If a bank stops spending on AI or mobile infrastructure, they risk losing their customers to tech-aggressive competitors; however, if they do spend, they aren't necessarily seeing their profit margins rise. This is the modern parallel of the Solow Paradox, based in the rapidly developing AI era. 

1. The "Dual-Staffing" Transition

My findings (or lack thereof) regarding Revenue per Employee (Model 5) highlight the primary hurdle to profitability: the Dual-Staffing Trap.

The Theory: AI should automate away routine tasks and lower headcount costs.

The Reality: Firms are currently in a "hybrid" phase. They must hire high-priced AI engineers and cloud architects to build the future networks, while simultaneously maintaining a legacy workforce to manage old systems and complex regulatory compliance. This "double-payroll" effect is what’s potentially keeping the J-Curve in its downward dip.

2. Is there an "AI Bubble"?

The aggressive spending by Fintechs (often exceeding 20% of their total expenses) mirrors the tech-heavy investment attitude of the late 90s. However, my evidence for the J-Curve (Model 4) offers a more optimistic outlook. The fact that the "profit penalty" shrinks by nearly 30% (from -0.848 to -0.594) when we look at lagged data suggests that the returns are coming—they are just moving at the speed of institutional change, not the speed of the stock market. This seems to contradict the theory of an AI bubble, however the question still remains: when will we see these returns?

3. Final Concluding Thoughts

After this extensive project, involving plenty of data analysis, my results act not as groundbreaking discovery, but as a confirmation to prior theories. However, the results of my 6 models remain important in the eye of an AI bubble. 

I did not disprove The Solow Paradox; it was simply updated for the new AI age. We see AI adoption everywhere, with companies racing to integrate it into every facet of their products. This process has been rapid, but my results prove that we won't see this reflected in firm profit margins until firms solve Dual-Staffing problems and move past the adoption phase of the J-Curve model.

Acknowledgements

1. Return of the Solow Paradox? IT, Productivity, and Employment in U.S. Manufacturing

2. Managing bank IT spending: Five questions for tech leaders

3.  The History of Large Language Models.

4. Bank IT Spending – Use These Metrics to Improve Performance