Intro to Finance & Economics, the conceptual ground beneath the math.

Many ML practitioners encountering finance for the first time wonder why it has its own vocabulary, its own benchmarks, its own conferences, and its own intellectual culture — surely a price-prediction problem is just supervised learning with a time index? The answer is that finance and economics aren't really about prediction. They are about how decentralised actors make decisions under scarcity and uncertainty, and prices are the emergent summary of those decisions. Once you internalise that frame, the methodology of the next chapter — and most of what financial firms do — makes much more sense.

Scarcity, trade, and the role of prices

The foundational economic problem is scarcity: there are more wants than resources, and someone has to decide who gets what. Markets are a particular institutional answer — let people trade with one another, voluntarily, and let the resulting prices coordinate who produces and who consumes. The Hayek-style insight is that prices aggregate distributed information: a rising price tells producers to produce more and consumers to consume less, without any central planner needing to know why demand has risen. This is why markets and prices are so fundamental to the discipline — they are the data structure through which decentralised information flows.

Finance is the part of economics that deals specifically with intertemporal choice — decisions that involve money or resources today versus tomorrow, with uncertainty about what tomorrow will bring. Borrowing, lending, investing, insuring, hedging, and speculating are all variations on the same underlying problem: how do we value claims on uncertain future cash flows, and how do we trade those claims efficiently? Almost every financial instrument in existence is some answer to that question.

Why "just predict the price" is the wrong frame

If you could reliably predict tomorrow's stock price, you would be very rich very quickly — and so would every other person who could. As soon as the prediction is widely known, traders act on it; their trades move the price toward the predicted level now; the prediction is incorporated into the current price; and the apparent edge disappears. This is the underlying reason markets are hard: prices contain almost all publicly available information by the time they're observed, and finding signal beyond that is what the entire industry of active management is competing for.

The implication for ML in finance is that the methodology must be more humble than in many other domains. The signal-to-noise ratio is small by construction — markets are constantly removing whatever signal becomes detectable. The chapter on financial ML develops the methodological discipline this requires; this chapter is about understanding why it's required.

Risk versus uncertainty

Frank Knight's 1921 distinction between risk (knowable probability distributions) and uncertainty (unknowable distributions, "unknown unknowns") is a cornerstone of how finance practitioners think. Most quantitative finance is about risk in Knight's sense — variance, value-at-risk, factor exposures, all computed on the assumption that the underlying distribution is stable enough to estimate. But financial markets are also subject to genuine uncertainty: unprecedented events (the 2008 crisis, COVID-19, currency regime changes) where the historical distribution simply doesn't apply. Mature financial-ML practice respects this distinction and uses different tools for each — statistical models for risk, scenario analysis and stress testing for uncertainty.

Positive versus normative

Economics distinguishes positive claims (descriptions of how the world is) from normative claims (statements about how it should be). "Raising interest rates reduces inflation" is positive. "The Fed should raise rates" is normative. Much economic disagreement is over normative questions where reasonable people with similar positive beliefs reach different conclusions because they weight the costs and benefits differently. Financial ML lives mostly in the positive domain — it tries to predict what will happen — but the deployment context (regulation, fairness, market manipulation) brings normative questions in immediately. The ML practitioner who pretends otherwise will be unprepared for the scrutiny their models receive.

A dollar today is worth more than a dollar tomorrow — partly because you could invest today's dollar and earn interest, partly because tomorrow's dollar might not arrive (default risk), partly because tomorrow's prices might be higher (inflation), and partly because humans simply prefer present consumption. The time value of money is the conceptual bedrock of finance, and almost every formula in the field is some elaboration of it.

Present value and future value

The fundamental conversion is between present value (PV) and future value (FV). If you invest $100 today at a 5% annual interest rate, after one year you have $105. After ten years, with annual compounding, you have $100 · (1.05)¹⁰ ≈ $162.89. The compounding works in reverse: a guaranteed $162.89 in ten years is equivalent to $100 today, at a 5% discount rate.

Net Present Value

Most real-world financial decisions involve sequences of cash flows — a project that costs $100 upfront and produces $20 a year for ten years; a bond that pays $5 a year for thirty years and returns $100 at the end; a mortgage that gives you $300,000 today in exchange for thirty years of monthly payments. The Net Present Value (NPV) of such a stream is the sum of the discounted cash flows:

Internal Rate of Return

The Internal Rate of Return (IRR) is the discount rate that makes the NPV exactly zero — it's the "break-even" rate of return. IRR is widely used in private-equity and project-evaluation contexts because it has a clean comparative interpretation ("this project earns 18% per year"), but it has well-known mathematical pathologies (multiple roots, scale insensitivity) that make it unreliable when comparing projects with different sizes or timing patterns. Modern corporate finance prefers NPV, but IRR remains the lingua franca of private-investment discussion.

Compounding and the rule of 72

An intuition pump: at a 7.2% annual return, money roughly doubles every 10 years. At 10% it doubles every 7.2 years. The general rule, known as the Rule of 72, is that the doubling time is approximately 72/r% years. This compounding effect is why small differences in returns matter enormously over long horizons — a 1% extra annual return over 40 years produces ~50% more terminal wealth, before any additional contributions. The tax and fee differences between investment products that look small in any single year are exactly this kind of compounding effect.

Why the discount rate matters more than the cash flows

A subtle but important point: in long-horizon valuation, the choice of discount rate often dominates the choice of cash-flow forecast. A discount rate of 5% versus 10% changes the present value of a 30-year cash flow by ~70%. This is why valuing infrastructure projects, pension liabilities, or technology companies (with cash flows decades in the future) is so contentious — the inputs that look most "objective" (the cash-flow projections) are often dominated by the input that looks most "subjective" (the discount rate). Financial ML that aims to value assets needs to engage with this carefully; pure return prediction can sidestep it but pure valuation cannot.

If all investments returned the risk-free rate, there would be no need for the rest of finance. The whole field exists because investments that promise higher returns also carry higher risk, and the central problem is how to characterise that trade-off, how to measure it, and how to choose among investments with different risk-return profiles. This section develops the conceptual vocabulary that everything in classical finance and most of financial ML builds on.

Expected return and variance

The simplest characterisation of an investment's risk-return profile uses the first two moments of its return distribution. The expected return is the probability-weighted average return across possible outcomes. The variance (or its square root, the standard deviation) measures how much returns disperse around the mean.

The Sharpe ratio

To compare investments with different risk levels, a single risk-adjusted metric is needed. The Sharpe ratio (William Sharpe, 1966) is the standard:

Diversification and correlation

The most important practical consequence of variance-based thinking is diversification: combining imperfectly-correlated investments produces a portfolio with less variance than any individual component. Concretely, if you hold equal weights of two stocks each with σ = 20%, the portfolio's σ depends on their correlation: ρ = 1 produces σ = 20%, ρ = 0 produces σ ≈ 14%, ρ = −1 produces σ = 0. The lower the correlation, the bigger the diversification benefit. This is the "free lunch" of finance — the only one most participants agree exists.

Diversification's power is bounded, however. Systematic risk (also called market risk) is the risk that affects all assets in the same direction — a recession, a financial crisis, a war. Diversification doesn't reduce systematic risk; only the idiosyncratic risk specific to individual assets gets diversified away. The remaining systematic component is what investors are compensated for bearing — the equity-risk premium, the bond-risk premium, etc.

The Capital Asset Pricing Model

The Capital Asset Pricing Model (CAPM, Sharpe-Lintner-Mossin 1960s) gives the first formal treatment of how systematic risk should be priced. The result: the expected excess return on any asset i is proportional to its beta — its sensitivity to the market portfolio.

CAPM is empirically false in its strong form — many anomalies (size, value, momentum, profitability) generate returns that beta alone cannot explain. But its conceptual structure — separating systematic from idiosyncratic risk, recognising that only systematic risk should be priced — is the foundation for the modern factor-model literature that the next chapter develops in detail.

Beyond mean-variance: tail risk

Variance is not the right measure of risk for many real-world purposes. Investors care more about losing 50% than about an equivalently-sized gain; they care more about a 1-in-100 catastrophic loss than about everyday wobble. Modern risk management uses metrics that capture the tail risk: Value at Risk (the loss that's exceeded only X% of the time), Expected Shortfall (the average loss conditional on being in the tail), and various drawdown-based metrics. The next chapter develops these in the financial-ML context; the conceptual point here is that "risk = variance" is a useful first approximation but a dangerous summary.

Where do prices come from? In a market with many participants, prices are the outcome of a continuous matching process — buyers and sellers placing orders, those orders matching, the resulting trades reported to the world. Understanding the mechanics of this process — supply and demand, equilibrium, the order book — is the precondition for understanding both the data financial ML works with and the structure of the next chapter's microstructure section.

Supply and demand

The textbook account: a demand curve shows how much buyers are willing to purchase at each price (typically downward-sloping — lower price, more demand); a supply curve shows how much sellers are willing to provide at each price (typically upward-sloping — higher price, more supply). The market clears at the price where the two curves intersect, and the resulting equilibrium price is what economists call the price.

This is an idealisation. Real markets are dynamic: shocks shift the curves, trades happen at non-equilibrium prices, and information is asymmetric. But the conceptual structure — price as the equilibrating mechanism between competing demands — is what makes markets work as resource-allocation mechanisms, and it underlies almost every financial-economics argument.

Auctions and continuous markets

The actual mechanism by which prices are formed varies. Auctions — like art auctions or treasury-bill auctions — collect bids over a period and clear once. Continuous markets, used for stocks and most modern liquid assets, run all day with new orders arriving continuously and matches happening as soon as compatible orders meet. The dominant continuous-market mechanism is the limit-order book: a sorted list of pending buy orders (bids) and sell orders (asks); a market order arriving on one side matches the best opposite-side limit orders immediately.

The structure of the limit-order book — how much volume sits at each price level, how quickly orders arrive and cancel — is the substrate of high-frequency trading and is the data on which microstructure ML operates. The next chapter's HFT section returns to this in detail.

Bid-ask spread and liquidity

At any moment, the highest bid price is below the lowest ask price — you can buy at the ask and sell at the bid, with a small gap (the bid-ask spread) representing the cost of immediate trading. Tight spreads (a few basis points) indicate liquid markets where lots of buyers and sellers are active. Wide spreads indicate illiquid markets where you pay more to trade quickly.

Liquidity is one of the most important concepts in finance. A liquid market is one where you can buy or sell substantial volume without moving the price much; an illiquid market is one where even small orders move prices significantly. The 2008 crisis was substantially a liquidity crisis: even high-quality assets became unsaleable because no buyers were willing to step in. The next chapter's execution-algorithms section is largely about minimising the cost of trading in markets with finite liquidity.

Primary versus secondary markets

A primary market is where new securities are issued — a company's IPO, a government bond auction, a new corporate-debt issuance. A secondary market is where existing securities are traded among investors — the New York Stock Exchange, the London Stock Exchange, the various electronic trading venues. Most of the financial-ML literature operates on secondary-market data; primary-market issuance is its own specialised sub-field.

Why "the market price" is more subtle than it sounds

News reports refer to "the price" of a stock as if there were a single number. There usually isn't. At any moment there's a bid, an ask, a last-traded price, and a mid-price; depending on the market and time of day there might be different prices on different exchanges. For financial-ML purposes, careful analysts spend considerable effort thinking about which of these prices is the right input for which question — using the bid for sell-side decisions, the ask for buy-side decisions, the mid for valuation, and the last-trade for accounting. This is the kind of thing that experienced quants take for granted but ML practitioners new to finance often miss.

If markets aggregate information efficiently, then the current price already reflects everything publicly known about an asset, and predicting future prices from public data is impossible. This is the Efficient Market Hypothesis, the most important and most-debated theoretical claim in finance. Anyone trying to apply ML to predicting markets is implicitly betting against some version of EMH, and understanding what they're betting against is essential.

The three forms

Eugene Fama's 1970 formulation distinguished three increasingly strong versions:

Weak-form efficiency: prices reflect all information contained in past prices. Implication: technical analysis (predicting future prices from past price patterns) cannot reliably make money. This form is well-supported by evidence — pure price-pattern strategies generally fail to outperform after costs.

Semi-strong-form efficiency: prices reflect all publicly available information, including financials, news, and analyst reports. Implication: fundamental analysis based on public data cannot reliably outperform. This form is more controversial — there are documented (small) anomalies that survive multiple-testing correction, but they are also small enough that they could be artifacts of model misspecification.

Strong-form efficiency: prices reflect all information, including private/insider information. Implication: even insider trading cannot make money. This form is rejected by evidence — insider trading clearly does generate excess returns, which is why it's regulated.

What EMH does and doesn't claim

EMH is widely misunderstood. It does not claim that prices are always "right" — they can be wrong, sometimes spectacularly. It claims that prices reflect all publicly available information, which is consistent with bubbles (everyone knew dot-com prices in 1999 were stretched, but many believed they would stretch further). It does not claim that markets are perfect — it claims that predicting them with public information is hard, not that they're efficient in any normative sense. And it does not claim that no one can outperform — it claims that doing so requires either luck, private information, or a genuinely better model than what's collectively priced in.

The Grossman-Stiglitz paradox

Sanford Grossman and Joseph Stiglitz (1980) pointed out that perfect EMH is logically self-contradictory: if prices already reflect all information, no one has incentive to spend money researching, so no one would gather the information that's supposedly already in prices. Markets must therefore be almost efficient — efficient enough that it's hard to make money, but inefficient enough that researchers can earn enough to compensate them for their effort. This is the conceptual underpinning of the modern active-management industry.

The implication for financial ML is the same: there must be exploitable signal somewhere, and the field's competitive structure means that signal is small and hard to find. The methodology of the next chapter — careful backtesting, multiple-testing correction, walk-forward analysis — is the discipline this requires.

Behavioural critiques

The 2000s and 2010s saw a substantial pushback against EMH from behavioural economics. Markets sometimes display bubbles, panics, and herding behaviour that strict EMH cannot easily explain. Limits to arbitrage — the noise-trader risk argument of De Long et al. (1990) and the slow-moving-capital arguments of more recent literature — explain why mispricings can persist even when sophisticated traders identify them. The 2026 working consensus is roughly: markets are usually close to efficient, but with predictable deviations during stress, and with sufficient effort, careful methodology can find small persistent edges.

What this means for ML practice

Three practical implications. First, most ML signals in finance are weak — Sharpe ratios of 0.3–1.0 are typical for individual signals, far below what's reachable in domains like image classification. Second, signals decay — once enough firms discover a signal, it gets arbitraged away; the next chapter's "alpha decay" framing captures this. Third, methodology matters more than model class — getting the cross-validation right, accounting for transaction costs, bias-correcting for multiple testing, all of these have larger effects on production performance than the choice between XGBoost and a transformer. EMH is not a wall the field gives up at; it is the frontier the field works to push back, slowly, with discipline.

A practitioner who has heard "stocks and bonds" but never looked at the broader landscape will be surprised at how many distinct asset classes exist, how different their statistical properties are, and how each comes with its own market structure, regulatory regime, and analytical conventions. This section is a tour of the major asset classes and the things an ML practitioner needs to know about each before applying methods to its data.

Equities (stocks)

Equity is a residual claim on a company's profits. Owning a share entitles you to a fraction of the company — voting rights, eventual dividends, and a share of any liquidation proceeds. Stocks are the most-traded and most-data-rich asset class. Returns are noisy at short horizons, with annualised volatilities of ~15–30% for individual stocks and ~15–20% for broad indices. Most quant-ML literature is on equities because the data is best.

Within equities, the dimensions of variation are: market capitalisation (large-cap, mid-cap, small-cap, with smaller stocks more volatile and less liquid); style (growth vs. value, with their own factor exposures); sector (technology, financials, healthcare, etc.); and geography (US, developed-international, emerging markets). Each combination has its own risk-return characteristics and its own ML methodology.

Fixed income (bonds)

A bond is a debt contract — the issuer borrows money and promises a fixed schedule of future payments (coupons plus principal at maturity). Government bonds (US Treasuries, German Bunds, Japanese JGBs) are typically the safest and most liquid; corporate bonds carry credit risk; municipal bonds in the US offer tax advantages; sovereign bonds from emerging markets carry both currency and default risk.

Bond markets are larger than equity markets globally (~$130T vs. ~$110T in 2024) but trade much less actively. Most bonds change hands rarely after issuance, and pricing is often quoted by dealers rather than by a continuous order book. The mathematical machinery for bonds (duration, convexity, yield curves) is its own substantial field, and term-structure modelling (Vasicek, Cox-Ingersoll-Ross, the various affine models) is a specialised quantitative discipline.

Foreign exchange (FX)

The FX market is the largest financial market in the world by trading volume — roughly $7T per day in 2024. Currency pairs (EUR/USD, USD/JPY, GBP/USD) trade 24 hours a day across global venues with extremely tight spreads and deep liquidity. FX is dominated by central-bank policy, balance-of-payments dynamics, and macro-economic news, and the ML methods used differ from equity-style factor models — the inputs are more macro-flavoured, the time scales are different, and the carry-trade structure (borrow in low-rate currency, lend in high-rate currency) is central to the literature.

Commodities

Oil, gold, copper, wheat, natural gas. Commodities trade primarily in futures markets (contracts for delivery at a future date) and have unique features: physical delivery, storage costs, seasonality (agriculture), supply shocks (weather, geopolitics), and structural producer-consumer relationships. The "convenience yield" — the benefit of holding the physical commodity rather than a paper claim — is a concept specific to this asset class. Commodity ML is its own specialised area, with substantial overlap with macro-economic forecasting.

Derivatives

Derivatives are contracts whose value derives from underlying assets. Forwards and futures are agreements to trade an asset at a specified future price. Options give the right (but not obligation) to buy or sell at a specified strike price. Swaps exchange one cash-flow stream for another (interest-rate swaps, currency swaps, credit-default swaps). Structured products combine these into custom payoffs.

Derivatives valuation has its own substantial mathematical apparatus — the Black-Scholes formula for European options, Monte Carlo simulation for path-dependent products, partial-differential-equation solvers, the various stochastic-volatility extensions. Modern ML for derivatives (covered briefly in the next chapter under "differential machine learning") replaces some of this machinery with neural-network approximators.

Crypto and digital assets

Bitcoin and Ethereum are the dominant cryptocurrencies, with thousands of smaller alternatives. Crypto is interesting from an ML perspective because the data is exceptionally rich — every transaction is on-chain and public, with timestamps, addresses, and amounts. It is also distinctive in being the most volatile major asset class (annualised volatilities of 50–100% for major coins) with weak fundamentals to anchor valuations. Crypto-specific ML methods include on-chain analysis (graph methods over the transaction network), DEX market making (automated market makers with their own mathematics), and smart-contract security analysis. The regulatory environment is unsettled, and serious quant deployment in crypto has been more cautious than the headlines suggest.

Real estate, private equity, hedge funds, and "alternatives"

Beyond the public markets, large pools of capital flow into alternatives: real estate (directly or via REITs), private equity (buyouts of private companies), venture capital (early-stage companies), hedge funds (active strategies on liquid markets), private credit (loans not traded on public markets), and infrastructure. These have weaker price-discovery mechanisms than public markets, distinct return distributions, and very different data — typically fund-level rather than asset-level. ML's role in alternatives is more limited but growing, particularly for sourcing investments and for due-diligence document analysis.

Asset prices come from markets, but the data those markets are pricing comes from the companies themselves — quarterly and annual filings that detail every dollar of revenue, every category of expense, every change in cash, and every entry on the balance sheet. Financial statements are the substrate of fundamental analysis, the source of the features that drive most equity-research pipelines, and the only language in which the financial health of a firm can be discussed precisely. An ML practitioner who cannot read a 10-K is missing the structured input that decades of financial-research methodology has been built on.

The three statements and how they connect

Every public company produces three primary financial statements. The income statement (also called the profit-and-loss statement, or P&L) reports revenue, expenses, and net income over a period (a quarter or a year). The balance sheet reports assets, liabilities, and shareholders' equity at a single point in time (the end of the period). The cash flow statement reports the actual movement of cash into and out of the business during the period, broken into operating, investing, and financing activities. Together they form a connected system: net income from the income statement flows into retained earnings on the balance sheet; the cash flow statement reconciles the change in cash on the balance sheet to net income via accruals adjustments. A statement that doesn't tie together is a sign of either accounting error or fraud.

The conceptual difference matters. The income statement is built on accrual accounting — revenue is recognised when earned, expenses when incurred, regardless of when cash actually changes hands. This produces a smoother and more economically meaningful picture of profitability than cash flows alone, but it gives management substantial discretion (when exactly was the revenue earned? when did the expense get incurred?). The cash flow statement is the corrective: cash is harder to manipulate than accruals, and large divergences between net income and operating cash flow are one of the classic warning signs of accounting mischief.

The income statement

The income statement reads top-to-bottom as a stack of subtractions starting from revenue. The structure of a typical statement:

Revenue (also called sales or top line) — the gross amount of business done in the period, before any costs. For a SaaS company, recurring subscription billings; for a retailer, total products sold times price; for a bank, interest plus fees earned. Modern revenue accounting under ASC 606 / IFRS 15 has detailed rules about when revenue can be recognised — particularly important for software, services, and contracts with long delivery cycles.

Minus cost of goods sold (COGS, also called cost of revenue) — the direct cost of producing what was sold. For a manufacturer, raw materials, factory wages, and depreciation of factory equipment; for a software firm, hosting costs and customer-facing engineers; for a retailer, the wholesale cost of merchandise. The remainder is gross profit; gross profit divided by revenue is gross margin, the most important measure of pricing power.

Minus operating expenses — the indirect costs of running the business, typically split into research and development (R&D), sales and marketing (S&M), and general and administrative (G&A). The remainder is operating income (also called EBIT — earnings before interest and taxes). EBIT divided by revenue is operating margin, the second most important profitability measure.

Minus interest expense, plus interest income, plus or minus other non-operating items. The remainder is pre-tax income. Minus taxes. The remainder is net income, the famous "bottom line." Net income divided by the share count is earnings per share (EPS), the metric most tracked by analysts and most quoted in news.

A subtle and increasingly important variant is EBITDA — earnings before interest, taxes, depreciation, and amortisation. EBITDA strips out the non-cash depreciation and amortisation charges to produce a measure closer to operating cash generation, useful for comparing companies with different capital intensities or capital structures. Critics (Buffett famously) call EBITDA "earnings before the costs that actually matter"; defenders point to its utility for relative valuation. Both are right; both EBITDA and net income should be looked at together.

The balance sheet

The balance sheet snapshots what the company owns and owes at a specific moment, structured as the accounting identity:

Assets are the resources the company controls. They split into current assets (expected to be converted to cash within 12 months — cash, marketable securities, accounts receivable, inventory) and non-current assets (longer-term — property, plant, and equipment (PP&E), intangibles like patents, goodwill from acquisitions, long-term investments). The split matters: a company with abundant non-current assets but little cash and large bills coming due may face a liquidity crisis even if it's profitable on paper.

Liabilities are the company's obligations. They split similarly into current liabilities (accounts payable, short-term debt, accrued expenses, the current portion of long-term debt) and non-current liabilities (long-term debt, pension obligations, deferred tax liabilities). The maturity profile of debt — when each tranche comes due — is one of the most important things to know about a company's risk profile, and is rarely visible from a single ratio.

Shareholders' equity is the residual — assets minus liabilities. It includes contributed capital (what shareholders paid for new stock issued, plus the par value of common stock), retained earnings (accumulated net income from all prior periods, minus dividends paid out), and various adjustments (treasury stock from buybacks, accumulated other comprehensive income). The book value of equity is the accounting measure of what shareholders own; it differs from market value (the stock price times share count) in ways that drive the value-vs-growth distinction in factor investing.

The cash flow statement

The cash flow statement starts from net income (the bottom of the income statement) and reconciles it to the change in cash on the balance sheet. The structure splits cash flows into three categories:

Cash from operating activities (CFO): the cash actually generated by the underlying business. Starts with net income; adds back non-cash charges (depreciation, amortisation, stock-based compensation); adjusts for changes in working capital (a buildup of receivables or inventory consumes cash; a buildup of payables generates cash). Strong CFO is the single most important sign of business health — it's the cash available to pay back debt, pay dividends, buy back stock, or reinvest. Companies whose CFO consistently lags reported net income (high accruals) are flagged as accounting-quality risks.

Cash from investing activities (CFI): cash spent on long-term assets (capital expenditures, acquisitions, investments) minus cash from divestitures. CFI is typically negative for growing companies (they're investing in PP&E) and positive for liquidating ones. The ratio of capital expenditures to depreciation is a useful measure of growth investment — well above 1 means the company is expanding its asset base, well below 1 means it's running down its asset base.

Cash from financing activities (CFF): cash from issuing or buying back stock, plus debt issued minus debt repaid, minus dividends paid. CFF tells you whether the company is returning capital to shareholders (negative — buybacks and dividends) or raising it (positive — issuing stock or borrowing). Mature companies typically show negative CFF as they distribute profits; growth companies often show positive CFF as they raise capital.

The sum of CFO + CFI + CFF equals the change in cash on the balance sheet. Free cash flow (FCF) is the most-watched derived metric: typically defined as CFO minus capital expenditures, FCF measures the cash a business generates after maintaining its asset base. FCF is the input to discounted-cash-flow valuations and the bottom-line summary of cash productivity.

Key ratios and what they tell you

Raw line items become useful when normalised into ratios. The standard families:

Profitability: gross margin (gross profit / revenue), operating margin (EBIT / revenue), net margin (net income / revenue), return on equity (ROE = net income / shareholders' equity), return on assets (ROA = net income / total assets), return on invested capital (ROIC = NOPAT / invested capital). ROE and ROIC are the most important profitability measures — they capture how much profit the business generates per unit of capital deployed. Sustained ROE above the cost of capital is what separates value-creating companies from value-destroying ones.

Liquidity: current ratio (current assets / current liabilities, ideally > 1), quick ratio (current assets minus inventory / current liabilities, a tighter measure that excludes inventory which may not be sellable quickly), cash ratio (cash plus marketable securities / current liabilities). Liquidity ratios measure the company's ability to pay near-term obligations and are particularly important during stressed conditions when external financing dries up.

Leverage: debt-to-equity ratio (total debt / shareholders' equity), debt-to-EBITDA (often used for credit analysis), interest coverage (EBIT / interest expense, measures the cushion between earnings and debt service). Highly leveraged companies have higher returns on equity in good times but face existential risk in bad times — leverage amplifies returns in both directions.

Valuation: price-to-earnings ratio (P/E = stock price / EPS), price-to-book (P/B = market cap / book value of equity), enterprise-value-to-EBITDA (EV/EBITDA, where EV = market cap + debt − cash). These ratios connect financial-statement data to market prices, telling you what the market is paying for each dollar of earnings, book value, or cash flow. Low values indicate "value" stocks; high values indicate "growth" stocks; both are loadings on the Fama-French value factor of Section 3 (the value factor's systematic outperformance is the canonical anomaly that motivated the field).

Reading red flags

Financial statements can mislead, and the practitioner who reads them carefully knows what to watch for. Common warning signs:

Accruals divergence: net income consistently exceeds operating cash flow by large amounts. Accruals are the difference between accrual-basis profit and cash-basis profit; persistent high accruals suggest the company is recognising revenue earlier than cash arrives, which can be legitimate (long contracts) but is also the classic shape of revenue manipulation. The Sloan accruals anomaly (Richard Sloan, 1996) — that high-accrual stocks systematically underperform low-accrual stocks — is one of the most-replicated findings in accounting research.

Cookie jars: large reserves established in good periods and quietly released in bad periods to smooth reported earnings. Aggressive earnings smoothing is detectable as too-low variance in reported income relative to underlying business volatility.

Channel stuffing: shipping product to distributors aggressively at quarter-end to book revenue, even though the distributors don't actually need the product. Often visible as a buildup in receivables that exceeds revenue growth.

Round-number patterns: reported earnings unusually often hitting analyst consensus or just clearing a round threshold. Benford's-law-style analyses of trailing-digit distributions in reported numbers can flag suspicious patterns.

Aggressive non-GAAP adjustments: many companies report a "non-GAAP" or "adjusted" earnings number alongside the official GAAP figure. The adjustments are sometimes legitimate (excluding one-time legal settlements) and sometimes suspect (excluding "non-recurring" charges that recur every year). Comparing GAAP and non-GAAP earnings and looking at what's being excluded is a standard analyst practice.

Auditor changes and restatements: a company that switches auditors or restates prior financials should attract more scrutiny. Restatements are common enough to be unremarkable individually, but the frequency and magnitude matter.

Where the data comes from

US public companies file financial statements with the Securities and Exchange Commission via standardised forms. The 10-K is the annual report (audited, filed within 60–90 days of year-end). The 10-Q is the quarterly report (unaudited, filed within 40–45 days of quarter-end). The S-1 is the IPO prospectus. Companies in financial distress file 8-K disclosures of material events between regular reports. The full text of these filings is available free on the SEC's EDGAR system.

International equivalents have different names but similar structure. UK companies file annual reports under FRS standards or IFRS; EU companies follow IFRS via national securities regulators; Japanese companies follow J-GAAP or IFRS; emerging-market firms typically follow IFRS or local-equivalent standards. The major data vendors (S&P Capital IQ, FactSet, Bloomberg, Refinitiv) standardise these into common schemas that ML pipelines can consume.

How financial-ML uses statement data

The next chapter develops the methodology in detail; here is the conceptual map. Fundamental factors are the most direct application — the value, profitability, and quality factors of the Fama-French and follow-on literature are computed directly from financial-statement data. Earnings-surprise strategies trade on the difference between reported earnings and analyst expectations; the post-earnings-announcement drift (PEAD) is one of the strongest documented anomalies. Accruals quality models trade on the Sloan accruals anomaly directly. Financial-distress prediction uses balance-sheet and cash-flow features to forecast bankruptcy or credit downgrades. Sector and industry classification from text in the 10-K's MD&A section feeds into peer-relative valuation models. The 2024 generation of LLM-based financial NLP increasingly extracts structured features from the qualitative parts of filings — risk factors, management discussion, footnote disclosures — that classical financial-statement analysis treated as out of scope.

Markets do not exist in a vacuum — they sit on top of an economy whose growth, inflation, and policy regime drive most of what happens at the asset level. A practitioner who tries to model markets without understanding the macro substrate will be confused by regime changes that look mysterious only because the underlying drivers are invisible. This section is the macro background an ML practitioner needs to make sense of why financial data behaves the way it does.

GDP and the business cycle

Gross Domestic Product (GDP) is the total market value of goods and services produced in an economy in a period — the standard measure of economic output. GDP growth fluctuates over the business cycle: expansions (rising output, falling unemployment, rising inflation) followed by recessions (falling output, rising unemployment, possible deflation). The cycle is irregular but recurrent, with average expansions of 5–10 years and recessions of 6–18 months in modern US data.

Asset returns are correlated with the cycle. Equities tend to outperform during expansions and underperform during recessions; high-quality bonds tend to outperform during recessions (as central banks cut rates and investors flee to safety); commodities are cyclical with industrial demand. Cross-asset allocation models — the foundation of multi-asset investing — depend explicitly on understanding the cycle's effects.

Inflation

Inflation is the rate at which the general price level rises — the dollar-denominated cost of a representative basket of goods. Moderate inflation (2–4% annually) is considered healthy in most modern economies; high inflation (10%+) is destabilising; deflation (negative inflation) is feared because it can spiral (consumers delay purchases, demand falls, prices fall further).

Inflation matters for finance in several ways. Bond returns are mostly nominal — fixed dollar payments — so unexpected inflation erodes real returns. Equities are partly inflation-protected (firms can sometimes pass through cost increases) but the relationship is messy. Cash and short-term bonds lose purchasing power during inflation. The real interest rate (nominal rate minus inflation) is what matters for most economic decisions, and central banks target it indirectly through their nominal-rate policy.

Monetary policy

Central banks (the Federal Reserve in the US, the European Central Bank in Europe, the Bank of Japan, the People's Bank of China) set short-term interest rates and conduct asset purchases to influence economic conditions. Monetary policy is the dominant short-term driver of most asset prices — rate decisions move bond yields directly and equity prices indirectly via discount-rate effects.

Modern monetary policy has gone through several regime shifts. The 2008 crisis brought quantitative easing — central-bank purchases of long-dated assets to push down long-term rates when short-term rates were already at zero. The 2022 inflation surge brought rapid rate hikes. The 2024–2026 environment has been characterised by relatively higher real rates than the 2010s. Each regime change has produced substantial cross-asset return shifts, and any time-series ML model trained across these periods needs to handle the regime structure explicitly or it will be biased toward whichever regime dominates the training data.

Fiscal policy

Government spending and taxation — fiscal policy — also affect markets, though with longer time scales and more political contingency than monetary policy. Sustained deficits push up bond yields (more supply); tax-policy changes shift corporate profits; major spending programs shift sector returns. Fiscal policy is less directly modellable than monetary policy because it depends on political processes that are harder to forecast, but its first-order effects on bond markets and tax-sensitive sectors are real.

Exchange rates and the global system

Floating exchange rates mean that monetary-policy divergences across countries produce currency moves. A central bank that hikes rates more aggressively typically sees its currency appreciate (capital flows toward higher yields). This creates feedback effects across markets — a strong dollar hurts US exporters and emerging-market borrowers with dollar-denominated debt; a strong yen hurts Japanese exporters. The global financial system is interconnected enough that local shocks propagate quickly, and any cross-asset ML model needs to incorporate these dynamics.

Why macro matters for financial ML

Most published financial-ML papers ignore macro, focusing on cross-sectional or short-horizon relationships within asset classes. This works as long as the macro environment is stable, but produces brittle models that break during regime changes. Mature production financial ML at the largest funds explicitly conditions on macro state — different models for high-rate vs. low-rate environments, regime-switching architectures, macro features as inputs to cross-sectional models. The next chapter mentions some of these techniques but is primarily about within-regime methodology; the macro backdrop is what makes regime-aware methodology necessary.

Classical finance is built on the assumption that market participants are rational expected-utility maximisers. They aren't, and the deviations are systematic enough that an entire field — behavioural economics — has documented how. This section covers the most-important behavioural patterns, why they matter for markets, and how they shape both the persistence of certain anomalies and the regulatory framework that governs financial markets.

Heuristics and biases

Daniel Kahneman and Amos Tversky's program of research, starting in the 1970s, documented systematic deviations from rational decision-making. The major ones:

Loss aversion: people feel losses more strongly than equivalently-sized gains (roughly 2× as strongly in typical experiments). Implication for markets: investors hold losers too long ("disposition effect") and sell winners too early.

Anchoring: estimates are biased toward whatever number was mentioned first, even if irrelevant. Implication: stock-price anchoring on round numbers, IPO prices, recent highs.

Availability bias: events that come to mind easily (because they were recent or dramatic) are over-weighted in probability estimates. Implication: investors over-react to recent news and under-react to long-term changes.

Overconfidence: most people rate their abilities above the median, including in domains where they have no special skill. Implication: traders trade too much, mutual-fund managers under-perform passive benchmarks.

Confirmation bias: people seek and over-weight information that confirms existing beliefs. Implication: investors rationalise their existing positions rather than re-evaluating based on new information.

Prospect theory

Kahneman and Tversky's prospect theory (1979) replaces classical expected-utility theory with a more empirically-accurate model of how people actually evaluate risky choices. The key elements: people evaluate outcomes relative to a reference point (typically the status quo); they are risk-averse over gains but risk-seeking over losses; small probabilities are over-weighted, large probabilities under-weighted.

Prospect theory explains several real-world phenomena that expected-utility cannot. People buy lottery tickets (over-weighted small probability of large gain) and insurance (over-weighted small probability of large loss) simultaneously, which is incoherent under expected utility but coherent under prospect theory. Investors gamble after losses to break even (risk-seeking over losses) and lock in gains too quickly (risk-averse over gains). The implications for market dynamics — particularly for the equity-risk premium and the disposition effect — have been substantial.

Limits to arbitrage

Even if some traders make systematic mistakes, classical theory says rational arbitrageurs should profit by trading against them, eventually pushing prices toward fair value. The limits to arbitrage literature (Shleifer and Vishny 1997, De Long et al. 1990) explains why this doesn't always work. Arbitrageurs face their own constraints — capital limits, career risk, noise-trader risk (the mispricing might get worse before correcting and the arbitrageur's investors might pull funds), funding costs, leverage constraints. These frictions allow mispricings to persist longer than naive theory predicts.

The implication for financial ML: the existence of an "anomaly" in the data is not always evidence of a model error or a true alpha — it might simply be a mispricing that's too costly to arbitrage at scale. Mature quant practice distinguishes between exploitable mispricings and persistent-but-unprofitable ones.

Bubbles and crashes

Periodically, asset prices rise to levels that are difficult to justify on any fundamental basis (the dot-com bubble, the 2007 housing bubble, the 2017 and 2021 crypto bubbles), then crash. Classical theory has trouble accommodating bubbles; behavioural finance treats them as natural consequences of overconfidence, herding, and limits to arbitrage. Robert Shiller's work on this earned him the 2013 Nobel Prize in Economics; his "irrational exuberance" framing remains the standard.

For financial ML, bubbles and crashes are the regimes in which classical models fail most spectacularly. A model trained mostly on calm-market data will under-estimate tail risk; a model that doesn't account for the possibility of regime changes will be miscalibrated during transitions. Stress-testing under hypothetical crisis scenarios is part of mature risk-management practice for exactly this reason.

What this means for the next chapter

The classical theory the next chapter draws on (factor models, mean-variance optimisation, risk-neutral pricing) is built on rationality assumptions. The methodology of financial ML is robust enough to work with imperfectly rational markets, but the practitioner who understands the behavioural underpinnings will build better models, recognise the regimes in which classical assumptions fail, and not be surprised when markets do things that "shouldn't" happen.

The institutional substrate of finance — the banks, insurance companies, asset managers, regulators, and central banks that make markets work — is often invisible to the ML practitioner who is just modelling price data. But this institutional layer determines what data exists, who can access it, what regulations constrain models, and what risks exist that pure return-prediction misses. A working understanding of financial intermediation is essential for navigating production deployments.

Banks and the credit system

Commercial banks take deposits and make loans. The economically-essential function is maturity transformation — borrowing short (deposits, which can be withdrawn at any time) and lending long (mortgages, business loans, consumer credit). This transformation creates value (savers want liquidity, borrowers want long-term funding) but creates fragility (a bank run happens when deposits are withdrawn faster than long loans can be liquidated).

The 2008 financial crisis was substantially a crisis of shadow banking — non-bank entities (money-market funds, securitisation vehicles, repo markets) that performed bank-like maturity transformation without bank-like regulation. The post-crisis regulatory framework (Dodd-Frank in the US, Basel III globally) substantially restricts what banks can do; the residual financial-stability risk has migrated into adjacent institutions that ML practitioners should be aware of.

Insurance companies and pension funds

Insurance companies and pension funds are the largest pools of long-term capital in most developed economies. They have liabilities matched to specific future cash flows (insurance payouts, retirement benefits) and invest to fund those liabilities. Their investment behaviour differs from individual investors — they are buy-and-hold, are long-horizon, and are typically forced sellers only when liability shocks occur. ML for these institutions focuses on liability matching, ALM (asset-liability management), and long-horizon return forecasting.

Asset managers and the active-passive split

Asset managers manage capital on behalf of others — mutual funds, hedge funds, private-equity firms, sovereign-wealth funds. The split between active management (trying to outperform a benchmark) and passive management (matching a benchmark cheaply) has shifted dramatically over the last 20 years. Vanguard, BlackRock's iShares, and similar passive providers now manage roughly half of US equity assets, up from negligible levels in the 1990s. The shift has implications for market structure — passive flows produce predictable buying and selling patterns that ML strategies can exploit.

Hedge funds are a smaller but more economically interesting segment — typically high-fee, high-alpha-seeking, often quantitative. The largest quantitative hedge funds (Renaissance Technologies, Two Sigma, DE Shaw, Citadel) employ thousands of researchers and run trading strategies at scale that span the methodology of the next chapter and beyond.

Central banks and monetary stability

Central banks — the Fed, ECB, BoJ, PBoC, BoE — play several roles. They are the lender of last resort to commercial banks during crises (preventing bank runs from cascading). They set monetary policy by manipulating short-term interest rates and by buying/selling assets to influence longer-term rates. They regulate the banking system. They issue the official currency. Modern central banks operate with substantial independence from elected governments, on the theory that monetary policy should be insulated from political short-termism.

For financial ML, central-bank decisions are among the most-watched events. Policy meetings produce immediate market reactions; policy regime changes produce longer-term shifts in correlations and volatility. A substantial sub-industry of "Fed-watching" — analysts and ML systems that try to forecast central-bank decisions or react quickly to announcements — sits at the intersection of macro forecasting and short-horizon trading.

Payments and market plumbing

The plumbing of the financial system — clearing, settlement, custody, payments — is largely invisible but enormously consequential. The DTCC clears US equities; CHIPS and Fedwire process US dollar payments; SWIFT routes interbank messaging; Euroclear and Clearstream settle European bonds. Each piece of infrastructure has its own operational risks, regulatory regime, and relevant data. ML applications in this domain — fraud detection, AML, settlement-risk modelling — are substantial and growing.

Regulation as a first-class concern

Finance is one of the most heavily regulated industries on earth. Banks are regulated by central banks and dedicated banking authorities (the OCC and FDIC in the US, the EBA in Europe). Asset managers are regulated by securities authorities (the SEC in the US, ESMA in Europe). Insurers are regulated by state-level authorities in the US and the European Insurance Authority in Europe. Markets themselves are regulated to prevent manipulation and to protect investors. Cross-cutting frameworks (Dodd-Frank, MiFID II, the EU AI Act's high-risk financial provisions) add yet more constraints. Production financial ML cannot ignore this — every model deployed in a regulated context must satisfy explainability, fairness, and audit requirements that don't exist in most other ML domains.

This chapter has covered the conceptual foundations of finance and economics that an ML practitioner needs before encountering the methodology of the next chapter. This final section briefly surveys where ML enters this picture, points to which earlier sections matter for which next-chapter material, and provides a pointer to the more comprehensive references for deeper study.

Where ML enters finance

Modern financial firms use ML in roughly five categories of application. First, alpha research — predicting which assets will outperform — sits on top of the factor-model framework of Section 3 and the EMH discussion of Section 5. Second, risk modelling uses the variance, correlation, and tail-risk concepts of Section 3 in heavier mathematical machinery. Third, execution applies optimisation and reinforcement learning to the market-microstructure structures of Section 4. Fourth, fraud and credit decisioning applies classification and graph methods to the institutional context of Section 9. Fifth, macro and asset-class allocation draws on the macroeconomic background of Section 7.

The next chapter (Financial ML & Quantitative Methods) develops the methodology for each of these. The conceptual foundation here is the prerequisite for understanding why each method does what it does.

What each section enables in the next chapter

For navigation: the time-value-of-money material of Section 2 is implicit in any return-based model; the risk-return material of Section 3 is the foundation of Section 5 (Portfolio Construction) of the next chapter; the markets-and-microstructure material of Section 4 is the foundation of Sections 6–7 of the next chapter (HFT and Execution); the EMH material of Section 5 is the conceptual backdrop for the entire alpha-research discussion (Sections 2–4 of the next chapter); the asset-class material of Section 6 is the substrate on which all the methodology operates; the financial-statements material of Section 7 is the source data for fundamental factors (Sections 2–3 of the next chapter); the macro material of Section 8 informs regime-aware modelling; the behavioural material of Section 9 explains why naive rationality assumptions fail and where the predictable mispricings come from; the institutional material of Section 10 is the context for fraud, credit, and regulatory considerations (Sections 8–10 of the next chapter).

What this chapter doesn't cover

This is an introductory chapter, deliberately compact. It doesn't develop the mathematical apparatus of finance — option pricing, term-structure modelling, stochastic calculus — that quantitative finance practitioners need. It doesn't cover the institutional history of major events (the Great Depression, Bretton Woods, the 2008 crisis) that shaped current practice. It doesn't develop accounting and corporate finance — the substrate of fundamental analysis. Each of these is a subject of substantial textbooks; the references below point to the major ones for readers who want to go deeper.

The next chapter, in context

With this background in place, the next chapter (Financial ML & Quantitative Methods) develops the methodology of applying ML to financial problems. The reader who has internalised this chapter will recognise the conceptual structure of the next: alpha research builds on factor models; backtesting discipline addresses the noise-and-multiple-testing problem that EMH implies; portfolio construction builds on mean-variance and the risk-return trade-off; high-frequency trading operates on the microstructure of Section 4; execution algorithms tackle the market-impact problem implicit in finite liquidity; fraud, credit, and AML inhabit the institutional context of Section 9. The methodology will feel less mysterious because the problem will feel familiar.

Frontier and where finance is going

The frontier of finance and economics in 2026 includes several active areas. Climate-risk pricing: how to value assets exposed to physical and transition climate risks. Crypto and decentralised finance: a new institutional layer with its own market structure, regulatory uncertainty, and analytical methodology. Macroprudential regulation: post-2008 frameworks for managing system-wide financial risk rather than individual-firm risk. AI in financial services: the transformation of customer-facing financial services by LLMs, with substantial regulatory questions about disclosure, fairness, and model explainability. Behavioural-finance integration: the increasing acceptance that classical-rationality assumptions need supplementation in production models.

The financial industry of 2030 will look different from that of 2020 in ways that the ML practitioner working in finance should track. The conceptual tools of this chapter — markets, prices, time-value, risk-return, EMH, behavioural deviations, institutional substrate — will continue to be the substrate of how finance is talked about, even as the specific methodology evolves rapidly.