AI in Finance: A Gentle Introduction

Have you ever wondered how your bank knows how to block suspicious transactions before you even notice them? Or how investment apps create personalized portfolios that seem to read your mind? Here is what happened to millions of people just this week: loans approved in 30 seconds, investment recommendations delivered in-app within seconds, and suspicious transactions flagged (and sometimes blocked) in real time. This isn’t science fiction—it’s your financial life in 2026, powered by Artificial Intelligence (AI). And the most interesting part is that it’s showing up not only in Wall Street trading floors, but in everyday moments—when you tap to pay, apply for credit, or ask your banking app a simple question. 

How does AI make that possible? Think of AI as a financial detective that can spot patterns humans might miss. In finance, AI is like a super-smart assistant that never sleeps: it can process millions of data points in seconds and learn from every interaction. To understand where it shows up in banking and markets, it helps to know the core building blocks. 

AI systems in finance are built on key technologies, including Machine Learning (ML)—models that learn patterns from historical data and improve when retrained with new data. Just like you learn to recognize your favorite coffee shop’s busy hours, machine-learning systems learn patterns in spending, market trends, and suspicious activities; Natural Language Processing (NLP)—technology that works with human language (text and sometimes speech), enabling you to chat with your bank’s support bot or ask your financial app questions in plain English; Computer Vision—technology that interprets images, used in finance for tasks like scanning an ID during account opening or reading a pay stub to verify income; Reinforcement Learning—systems that learn by trial and error, for example improving trade execution over time by testing different order-splitting strategies to reduce transaction costs and price slippage (usually within strict risk and compliance constraints); Generative AI—systems that generate text (and sometimes code or summaries), for example drafting a plain-language explanation of a long bank statement or earnings report (with human or rule-based checks for accuracy); and Predictive Analytics—systems that forecast future outcomes as probabilities, like a weather report for your finances, estimating things like cash-flow needs, churn risk, or the likelihood of delinquency (and, in some contexts, short-horizon market signals).  

Why is this happening now? Simple: we generate more financial data than ever; customers expect instant responses, and everyone wants services tailored to them. AI is one of the few technologies that can deliver speed, scale, and personalization all at once—especially in an industry where small decisions compound into big outcomes. With that foundation in mind, let us zoom in on what it looks like when these models meet real-life banking: the everyday path your money takes from spending, to payments, to credit to investing. 

AI in Finance, in Everyday Life 

You might think AI in finance sounds futuristic, but chances are you have already experienced it today. In fact, many of the most important AI systems are built into routine moments that feel almost boring—until something goes wrong, and you realize how much is happening in the background. A helpful way to make it concrete is to follow a “day in the life” of your money—from spending, to borrowing, to investing—and notice where algorithms are quietly doing the heavy lifting. 

To keep it practical, I am focusing on the day-to-day path most people take where AI is already embedded in consumer experiences. 

Personal Banking 

Start with something simple: your morning coffee. When you buy it and immediately see the purchase categorized as “Food & Dining” in your banking app, that’s AI at work. The system recognizes the merchant and assigns the right spending category without any manual input from you. Sometimes it goes a step further and sends a gentle nudge: “You’ve spent $47 on coffee this week—15% more than usual.” Behind the scenes, machine-learning models analyze your transactions, compare them with your past behavior, and learn patterns in how and where you spend money. Over time, they get better at understanding your routines, so apps can offer personalized insights—suggesting budget adjustments, highlighting recurring expenses, or alerting you to unusual spending. In other words, AI turns raw transaction data into guidance you can actually use, helping you stay informed and in control without lifting a finger. 

Payments & Money Movement 

Right after you check your balance and buy that coffee, your money has to move—through card networks, digital wallets, and bank transfers. AI helps make those payments feel instant and “invisible” by scoring risk in real time: is this transaction typical for you, on this device, in this location, at this time of day? That risk score is essentially a fast probability estimate—how likely is this payment to be legitimate—computed in milliseconds using a mix of fraud rules and machine-learning models trained on past transaction patterns. If the pattern looks normal, the payment is approved quickly; if it looks off, the system might ask for extra verification (like a one-time passcode) or decline it before letting it through. The same logic shows up in person-to-person transfers and bill payments, where AI helps detect unusual activity and reduce errors—so you can send money quickly without giving up security. 

Compliance + Identity 

Another place you’ll encounter AI—often without realizing it—is during account opening and security checks. When a bank asks you to upload an ID, verify your identity, or answer “just a few more questions,” that’s part of Know Your Customer (KYC) and Anti–Money Laundering (AML) compliance. AI can help match documents, spot potential signs of identity theft, and flag transaction patterns that may indicate suspicious activity (like unusual transfer behavior or rapid movement of funds across accounts). Because these models can produce false positives, banks typically combine automation with human review and policy rules—especially for decisions like account closures or regulatory filings. For students and early-career professionals, it’s helpful to see this as the guardrail layer of modern finance: it protects customers, supports regulatory requirements, and builds trust—so when you move on to borrowing and credit decisions, the institution has more confidence in who it’s lending to and where the money is going. 

Lending and Credit 

With identity checks and compliance guardrails in place, the next step for many customers is borrowing—and that’s where AI plays an even more visible role. In fact, the same pattern-recognition power used to keep accounts secure also shows up in lending decisions. Remember when loan applications took weeks of paperwork and waiting? Today, many lenders can give an initial decision in minutes—and AI is a big reason why. Instead of looking only at a credit score, modern underwriting models may evaluate a broader picture: income and employment stability, existing debt obligations, cash-flow patterns, and consistency and completeness in the information you provide (the exact data used varies by lender, product, and regulation). These models don’t “approve loans on their own”—they produce risk estimates that lenders combine with policy rules, pricing, and fair-lending controls. For early-career borrowers, that broader view can make the process feel more transparent—because it reflects real financial behavior, not just a single number. 

This matters most for people with limited credit history. Rather than automatically rejecting someone with a “thin file,” some lenders can estimate creditworthiness using permissioned alternative signals—such as verified bank-account cash-flow data and, where available, rent or utility payment history—helping widen access to credit while still managing risk. For example, a consistent record of on-time rent and stable deposits may provide evidence of reliability even when a traditional credit score is still developing. Once a loan is approved, those same data-driven models may continue working in the background—monitoring for early warning signs of missed payments and helping lenders offer proactive support (like payment reminders or modified due dates) before a small problem becomes a default. 

Investment Management 

Once spending is categorized, payments are secured, and credit decisions are made, the next question becomes: what should you do with the money you’re trying to grow? After spending and borrowing, investing is the third place many people feel AI most directly—especially through apps designed for speed and simplicity. When an investment platform suggests rebalancing your portfolio, AI is working in the background—analyzing market conditions, your risk tolerance, and your financial goals at the same time. These systems, often called robo-advisors, differ from human advisors because they can evaluate data continuously and apply consistent rules without delays. In practice, a robo-advisor usually combines your inputs (goals, time horizon, and constraints) with portfolio models that turn them into an allocation plan, then recommends trades that keep you close to that plan as markets move. 

Rebalancing is only one piece of the puzzle. Beyond traditional price and fundamentals data, some platforms also process large volumes of unstructured information—such as financial news and analyst commentary—to gauge market sentiment and adjust recommendations accordingly. (Social-media signals exist too, but they tend to be noisy and are used more selectively depending on the firm and strategy.) For instance, a spike in negative headlines around a sector may prompt a more cautious allocation, depending on the strategy. Many robo-advisors can also execute trades automatically when preset conditions are met and update suggested portfolios as new information becomes available. 

In more advanced setups, AI-driven portfolio management can monitor your holdings, automatically reinvest dividends, and may harvest tax losses in taxable accounts—actions that can improve after-tax returns when done carefully (for example, by tracking tax lots and avoiding wash-sale violations). For finance students and early-career analysts, it’s worth noticing the underlying theme: AI is turning ongoing portfolio tasks into repeatable, rules-based workflows—making investing more scalable, while still leaving room for human judgment on goals, constraints, and ethics. 

Summary 

AI in finance isn’t one single tool—it’s a set of technologies powering the everyday journey of your money. The benefits are practical: faster decisions, more personalized insights, stronger risk controls, and scalable service. However, AI isn’t the solution to everything—human oversight is still essential to keep these systems accurate, fair, and meaningful. 

Now that we’ve seen the day-to-day use cases, the second blog will zoom out to cover the biggest trends shaping AI in finance, along with the challenges that come with them, including data privacy, model risk, bias and fairness, explainability, regulation, and where human oversight still matters most. 

The Stock Market Isn’t a Bell Curve (and Your Portfolio Shouldn’t Pretend It Is)

This year’s market swings made me step back and think about something we usually take for granted: the models and tools we use to make sense of it all. When markets get choppy, you start asking yourself—are these models actually capturing what is happening, or just giving me a comforting story? 

That’s when I realized something important: a lot of the tools investors rely on quietly assume markets behave in a “normal,” bell-curve way. It’s built into risk metrics, forecasts, and even the way many people size their positions. 

But real markets don’t look normal. They jump, cluster, and produce outlier days far more often than a neat bell curve would predict—and that has real consequences for investors. 
 
So this post digs into that gap: what the bell curve assumption actually implies, what S&P 500 data actually shows, and why your portfolio shouldn’t pretend the market is a bell curve. 

A look at real S&P 500 daily returns, fat tails, and what “risk” actually means for investors. 

What if I told you the stock market has “impossible” days far more often than most people—and many risk models—assume? 

Not because markets are broken. Not because history is uniquely unlucky. It’s because the neat bell-curve story is a poor description of how markets actually move—especially when it matters most: in the extremes. 

With that framing, I’ll use real S&P 500 daily return data to show (1) what the return distribution looks like in the real world, (2) which common distributions fit it best, and (3) what that implies for practical risk decisions like position sizing and drawdown planning. 

How to read the charts: the histogram/KDE shows the empirical distribution (what actually happened). The smooth model curves are theoretical “guesses” laid on top. The most useful model is the one that tracks the empirical curve—especially in the tails, where risk lives. 

Figures (the “show, don’t tell” part) 

As you read, pause at each chart and ask one question: “If I had built my risk plan assuming a bell curve, would this picture surprise me?” The captions are written so you can follow the story directly from the visuals. 

 Figure 1. Empirical distribution overlaid with fitted PDFs (Normal, Student’s t, Laplace, GED, Skew Normal). The shaded empirical shape is the test; the colored curves are “guesses.” What to notice: Normal looks plausible in the middle, then fails where risk lives: in the tails. 

 

 

At first glance, most models do a reasonable job near the center of the distribution. Most trading days cluster tightly around zero, producing a steep central peak. That’s why the normal (bell curve) remains so appealing: if you focus on typical days, it doesn’t look obviously wrong. 

But risk doesn’t live in the center. The tails reveal the problem. Move away from zero and the empirical distribution decays much more slowly than the normal curve. In practical terms, large moves occur far more often than a bell curve would predict. That’s where the bell-curve story breaks: Normal can fit the “average” day and still miss probability mass in the shoulders and tails—the part that determines whether a risk plan survives a crisis. Student’s t and GED stay closer because they allow heavier tails. 

Figure 2. Log-density view (same curves, log y-axis). This makes tail differences impossible to ignore. What to notice: the empirical curve decays slowly; thin-tail models decay too fast and understate extreme-day frequency. 

 

 

On a log scale, small differences become decisive. The empirical curve decays more slowly than Normal, meaning extreme days occur orders of magnitude more often than a bell curve would predict. Models that fall too fast here don’t just understate risk—they erase it. 

 Figure 3. Q–Q: empirical quantiles vs Normal and vs Student’s t. If a model fits, points trace a straight line; tail bends mean the model can’t reproduce real extremes. What to notice: Student’s t stays closer in the ends than Normal. 

 

 

The Q–Q plot asks a simple question: if a model were right, would the data’s extremes look like the model’s extremes? In that sense, it’s a truth serum—the tails are where models confess. Normal’s tail bending is exactly what “fat tails” means in practice. Student’s t doesn’t make markets safe; it just stops pretending extremes are impossible. 

 Figure 4. Tail survival of |return| on log–log axes: P(|R| > x). This answers: “How often do we get a move bigger than X%?” What to notice: Normal falls below the empirical tail—meaning it predicts too few big days. 

 

This is the investor’s frequency chart: “How often do big days show up?” If Normal sits below the empirical tail, it’s undercounting large-move days—and that gap fuels many “I didn’t think this could happen” moments. Put differently, the empirical tail decays far more slowly than Normal predicts. Heavy-tailed models track that reality more closely, confirming that extreme moves are meaningfully more common than a bell-curve world allows. Other distributions still miss the most extreme events, but Student’s t and GED stay far closer to the empirical tail. 

 Figure 5. Left-tail zoom (loss days only), empirical vs fitted models. This focuses on the “I didn’t expect that” part of investing. What to notice: models that match the center can still miss the downside tail. 

 

Focusing on loss days makes the asymmetry tangible. Many models look acceptable in the center, then misprice the downside tail once you zoom in. That’s why thin-tail assumptions are dangerous in practice: if you use them to size positions, you’re most likely to be wrong exactly when you can least afford it. 

 

Figure 6. Right-tail zoom (big up days), empirical vs fitted models. Bull-market euphoria has a distribution too. What to notice: upside extremes exist, but the left tail is typically the portfolio killer because of leverage, withdrawals, and investor behavior. 

 

Big up days are real—and part of the same fat-tailed world. But from a risk standpoint, the right tail rarely forces action; the left tail does (through leverage, withdrawals, risk limits, and emotion). That asymmetry is why downside modeling matters more for survival. 

 Figure 7. Rolling volatility (or rolling distribution width) over time. A single “average volatility” hides the truth: markets switch regimes. What to notice: calm periods can lull risk-taking right before volatility clusters arrive. 

 

 

With that regime-switching backdrop, rolling volatility shows why “average volatility” is a trap. Calm regimes invite bigger bets; crisis regimes punish them. The point isn’t to time regimes perfectly—it’s to avoid a plan that only works in the calm regime. 

Figure 8. Model fit comparison (AIC/BIC ranking). This is the “scorecard” behind the visuals. What to notice: Student’s t and GED beat Normal, which aligns with what the tail charts show. 

 

 

This figure turns the visual evidence into numbers. AIC and BIC rank models by fit while penalizing unnecessary complexity. The result is unambiguous: the scorecard confirms what the tail plots already suggested—heavier-tailed models fit better. Normal ranks last, not because it fails everywhere, but because it fails where it matters most: in the extremes. 

That doesn’t mean Student’s t is the “true” model of markets. It’s simply a less misleading baseline than Normal among common parametric choices, because it admits heavier tails—and therefore a more realistic frequency of extreme days. 

Figure 9. Risk metrics illustration: VaR vs Expected Shortfall (ES). VaR marks a threshold; ES tells you the average loss past that threshold. What to notice: in fat-tailed data, ES tends to be meaningfully worse than what a bell-curve intuition would suggest. 

 

Figure 10. Risk metrics illustration: Daily VaR vs Expected Shortfall (ES) at 95% and 99% confidence levels. VaR marks a loss threshold, while ES captures the average loss once that threshold is breached. What to notice: as confidence increases, ES deteriorates much faster than VaR, revealing the severity of downside risk in fat‑tailed return distributions 

This figure highlights a subtle but critical distinction in risk measurement. Value at Risk (VaR) answers: “How bad is a loss on a bad day?” Expected Shortfall (ES) asks the more practical follow-up: “When that bad day happens, how bad is it on average?” VaR sets a threshold, but says little about what lies beyond it. ES addresses that by focusing on the average severity of tail losses, not just the cutoff. 

In fat-tailed data, ES widens the gap between what feels plausible and what’s historically typical. For investors, that gap is the point: Expected Shortfall often implies more severe losses than VaR, and in fat-tailed markets that severity is closer to what you actually need to be prepared for. 

 With those pictures in mind, we can move from charts to implications—what the data says, and what it changes for real-world investing decisions. 

1) The story we tell ourselves about risk (and why it’s comforting) 

If you ask a room full of investors to describe “a normal market day,” you’ll hear the same story: small moves most of the time, occasional bumps, and truly wild days that are rare enough to ignore. That mental picture is basically a bell curve. 

• What I did instead: I started with daily S&P 500 returns , then compared the empirical distribution to several standard distributions investors often assume—especially the normal “bell curve.” 

Quick prediction: Before you look at the numbers, what do you think is more common—days above +3% or days below −3%? (Most people guess “about the same.” The data doesn’t.) 

2) Core descriptive statistics (the “shape” of the return distribution)  

Here’s the market’s “personality” in a handful of numbers. These stats sound technical, but they translate into a human experience: most days feel boring… until they suddenly don’t. 

• Observations: 1,346 daily returns 

• Mean daily return: ~0.000408 (≈ 0.0408%) 

• Daily volatility (std dev): ~0.00895 (≈ 0.895%) 

• Skewness: about -0.81 → more/larger downside moves than upside moves 

• Excess kurtosis: about 7.4 → very fat tails (Normal would be 0) 

• Worst day: -6.16% on 2020‑03‑12 

• Best day: +4.35% on 2020‑04‑16 

 

One detail worth sitting with: the worst day in this sample (−6.16% on 2020‑03‑12) and one of the best days (+4.35% on 2020‑04‑16) happened within weeks of each other. That’s not a “smooth” world; it’s a world where outcomes cluster—and where risk feels very different depending on when you arrive. 

This is not close to Gaussian: the distribution is left-skewed and fat-tailed, exactly the condition where “Normal-based” risk metrics can mislead. This aligns with the broader finance literature and practitioner guidance that equity returns often show fat tails and skewness, and that normal assumptions can understate tail risk. 

Mini thought experiment: imagine you invested at the February peak and checked your account at the March trough. Would you have stuck to your plan—or changed it? That’s why “distribution talk” matters: it maps directly to decisions made in real time. 

3) Market regime insight: the 2020 crash dominates tail behavior 

Using the Close series, the maximum peak-to-trough drawdown in the sample is approximately: 

• Max drawdown: -36.96% 

• Peak date: 2020‑02‑18 

• Trough date: 2020‑03‑31 

This is a concrete example of why tails matter: volatility and extreme moves arrive in bursts, not evenly over time. That “volatility clustering + tail events” pattern is widely discussed in fat-tail diagnostics for S&P 500 returns. 

Story #1: the leverage trap (why fat tails feel personal). Imagine an investor who uses mild leverage because the last few months look “stable.” A −3% day feels like a nuisance—until it arrives in a cluster. In a fat-tailed world, the real danger isn’t one bad day; it’s several bad days close together, when prices gap and rules-based selling (margin requirements, risk limits, de-leveraging) forces action at the worst time. 

Story #2: the retirement problem (sequence-of-returns risk). Now imagine a retiree withdrawing from a portfolio. Two investors can share the same long-run average return, but if one takes large losses early, withdrawals lock in the damage. Fat tails matter because the left tail isn’t just “paper loss”—it can permanently alter the trajectory of a plan. 

4) Closing: the one risk mistake investors repeat 

The market doesn’t punish investors for not knowing the future. It punishes them for building a plan that only works in a bell‑curve world. Your data points to fat tails: long quiet stretches, then clusters of big moves—especially on the downside. 

A quick investor checklist (keep it simple): 

• If your plan assumes “a −5% day is basically impossible,” adjust your expectations—and your position sizes. 

• Prefer tail-aware thinking: pair volatility with tail metrics (like Expected Shortfall) or simple stress scenarios. 

• Have a “tail plan” (cash buffer, rebalancing rules, and a written response plan) before the next crisis window.