Part I · Returns and risk · Chapter 1G

Where an edge comes from

Risk premia, behavior, institutions and market plumbing

When practitioners say "edge", they usually mean one of four very different things. Mixing them is the single most common analytical error in trading literature. It leads directly to over-allocating to overlays, the tactical positions laid on top of a portfolio, and under-allocating to genuine premia, the payments for bearing risk. The taxonomy in this chapter is the most important conceptual tool in this book's treatment of edges.

1 · The four sources

Every documented edge in financial markets traces back to one of four sources, or to a combination of them:

1. Compensated risk premia: payment for bearing risk that cannot be diversified away.
2. Behavioral and cognitive frictions: payment for being on the right side of systematic mispricing.
3. Institutional frictions and constraints: payment for accommodating predictable flow driven by mandates.
4. Microstructure and flow: payment for providing liquidity at high frequencies.

The four differ in why they pay, how durably they pay and who can capture them. Take the equity risk premium (ERP), the extra return that stocks are expected to pay over safe assets. Set beside it an OPEX-week tilt, a change of positioning in the week of the monthly options expiration (OPEX). Treating these two as comparable allocations is wrong, because they are fundamentally different objects that need different sizing, different persistence assumptions and different kill criteria. A kill criterion is a rule, set in advance, for abandoning a strategy.

2 · Source 1: compensated risk premia

What the island pays for a dollar in each kind of year

Let $q_g$ be today's price of a dollar paid only in a good year, and $q_b$ the price of a dollar paid only in a bad year. The bond pays $100 in both kinds of year and the stock pays $120 and $60, so their prices give two equations:

$$\begin{aligned} 100\,q_g + 100\,q_b &= 90 \\ 120\,q_g + 60\,q_b &= 70 \end{aligned}$$ (1G.1)

Substituting $q_b = 0.9 - q_g$ from the first equation into the second gives $60\,q_g + 54 = 70$, so $q_g = 0.267$ and $q_b = 0.633$.

The two kinds of year are equally likely, yet the island pays 63 cents today for a bad-year dollar and only 27 cents for a good-year dollar. A dollar that arrives when you need it is worth more than twice one that arrives when you do not. The stock is cheap because it delivers most of its dollars in good years.

The figure keeps the island's prices for a dollar in each kind of year and lets you reshape the stock. Drag its bad-year payoff down and watch the premium grow while its expected payoff stays at $90.

bad-year payoff

$60

bad-year dollar costs

63.3¢

stock price = $70.00

expected return = 28.6%

risk premium = 17.5 pts

good-year dollar costs 26.7¢

The stock pays $120 in a good year and $60 in a bad one, $90 on average. At $70.00 it returns 71.4% in a good year and −14.3% in a bad one, 28.6% on average, against the bond's certain 11.1%.

For any compensated risk premium, the formal object is

where $R_i$ is the return on asset $i$, $r$ the risk-free rate, $\beta_i$ the asset's exposure to the priced risk factor and $\pi$ the price of that risk per unit of exposure. With a hypothetical $\pi = 5\%$, an asset with $\beta_i = 1.2$ is expected to earn $1.2 \times 5\% = 6\%$ above the risk-free rate.

The key feature is that the premium cannot be arbitraged away, because the risk is real. An arbitrage is a profit without risk, and none is on offer here. Whoever bears the risk is paid for bearing it.

What is in this category

• The equity risk premium: payment for risk that is correlated with consumption (Chapter 10B).
• The variance risk premium (VRP): payment for taking the other side of investors who want to hedge volatility (Chapter 10C).
• The term premium: payment for accepting interest rate risk on long bonds.
• Carry and the commodity convenience yield: payment for supplying storage or for accepting the risk that prices revert (Chapter 10E).
• Trend following, with caveats: partly compensation for accepting tail risk in exchange for diversifying behavior (Chapter 10D).

Why it matters: persistence

Risk premia are the most durable source of return. They do not disappear when more capital chases them, because the risk is the reason the premium exists. If more capital takes the risk, the premium might compress. It cannot go to zero: at zero, no one would bear the risk, and the market would break down. This is the strongest argument for anchoring a portfolio in risk premia, whose persistence is more robust than that of behavioral or institutional edges.

What it costs

Risk premia are not free lunches: in stress, they pay their cost. The ERP is paid for in drawdowns of 30% to 50%, which by the arithmetic of Chapter 1E need gains of 43% to 100% to repair. The VRP is paid for in crashes, when implied volatility explodes. Trend is paid for in long whipsaw periods, when prices keep reversing before a trend can form. The premium exists because of these costs.

A common retail confusion is to treat risk premia as if they were behavioral mispricings that should always pay, which they do not. They pay on average across regimes while hurting badly in specific regimes, and the average is what compounds.

3 · Source 2: behavioral and cognitive frictions

Investors systematically misprice information or extrapolate from it. Efficient-market theory holds that prices already reflect what is known, so mispricings should vanish quickly. They persist longer than the theory predicts, for two reasons. First, limits to arbitrage prevent rational traders from fully correcting them. Second, behavioral biases are remarkably stable over time: people do not become less prone to recency bias, the overweighting of recent events, just because they have read about it.

A behavioral edge typically takes the form of a pricing error $e_t$, the gap between what the market believes and the true value:

The error then decays at a predictable rate:

Here $\phi$ lies between 0 and 1, so that a fraction of the error persists into the next period, while $\varepsilon_t$ is noise with a mean of zero. That persistence, the AR(1) behavior of Chapter 3B, is what creates exploitable predictability. With a hypothetical $\phi = 0.6$, an error of 10 percentage points is expected to shrink to 6 in the next period and 3.6 in the one after. A trader who bets on the correction expects to earn 4 points in the first period and 2.4 in the second, before costs.

What is in this category

• Momentum, both cross-sectional and time-series: under-reaction followed by over-reaction. Cross-sectional momentum buys the recent winners among stocks and sells the recent losers. Time-series momentum is the trend following of Chapter 10D, which trades each asset on its own past.
• PEAD, post-earnings announcement drift: under-reaction to fundamental news.
• The value premium, in part: over-extrapolation of growth in glamour stocks, the expensive shares of fast-growing companies (Chapter 10F).
• Lottery demand and the low-volatility anomaly: over-payment for high-beta stocks with lottery-like payoffs (Chapter 10F).
• Seasonality effects, such as the Halloween and January effects: the mechanisms are less clear, but the patterns persist (Chapter 10G).

Why it matters: decay

Behavioral edges decay when capital exploits them. Once enough sophisticated investors learn about PEAD, they buy the post-earnings winners, and the drift compresses. McLean and Pontiff (2016) document a decline of 26% to 58% in factor returns after publication, depending on the factor. Jensen, Kelly and Pedersen (2023) document a decline of about 47% after publication across hundreds of factors.

The decay is not to zero, because institutional and behavioral frictions remain, but it is substantial. Do not assume that an edge found in 1990s data still works at the same magnitude.

What it costs

Behavioral edges typically have lower drawdowns than risk premia, since they are not insurance products. They can, however, have catastrophic single-event losses when the regime that supported them flips. Momentum lost 73% from March to May 2009. Value's drawdown from 2010 to 2020 was the worst in modern history. The costs are different from those of a risk premium, episodic rather than systematic, but they are real.

4 · Source 3: institutional frictions and constraints

Institutional frictions create predictable price pressure through mandates, leverage limits, accounting rules and regulatory constraints. Big institutional investors do not trade like the rational agents of economic theory. They trade according to mandates, the written rules that say what a fund must hold, which often produce predictable behavior. If you can identify the mandate-driven flow, you can position ahead of it.

Institutional frictions create flow imbalances, and the imbalances move prices:

Here $\Delta P_t$ is the price change over period $t$, $\text{Imbalance}_t$ the mandate-driven buying less the mandate-driven selling, $b$ the price move per unit of imbalance, $a$ a constant and $\varepsilon_t$ noise. Suppose, hypothetically, that $a = 0$ and each $1 billion of net forced selling moves the index down 0.1%. The pension fund's $168 million then moves it $0.168 \times 0.1\% \approx 0.017\%$, under 2 basis points. The premium is small per trade but mechanically repeating.

What is in this category

• Betting against beta (BAB): leverage-constrained investors bid up high-beta stocks, so low-beta stocks outperform (Chapter 10F).
• End-of-quarter rebalancing flows: 60/40 mandates create predictable flow between asset classes (Chapter 10G).
• Demand-based option pricing: asset managers' need for hedges pushes up put prices.
• Index reconstitution effects, largely dead: S&P 500 additions historically rallied (Chapter 10I).
• Forced liquidations and margin-call cascades, which create predictable reversals once stress passes (Chapter 1E).
• Securities lending and short-borrow markets: the "specials" premium, the high fee earned by lending out a stock that short sellers find hard to borrow.

Why it matters: persistence against regulatory risk

Institutional frictions persist as long as the constraint persists. They can disappear quickly when regulations or product structures change. The S&P 500 inclusion premium, the jump in a stock added to the index, was a real effect of about 7% in the 1990s. It has since fallen below 1% (Greenwood and Sammon 2024), because index funds got smarter about avoiding the obvious arbitrage.

This differs from risk premia, which persist as long as the risk is real, and from behavioral edges, which decay slowly. An institutional edge can vanish in a regulatory update.

What it costs

The costs are usually small per trade and concentrated in stress periods. When the flow gets dislocated, the institutional player you are trading against may default. Trade slippage and execution costs are the main retail enemies, because the edge per trade is small (Chapter 1F and Chapter 2D).

5 · Source 4: microstructure and flow

Microstructure edges are short-horizon predictability driven by order-flow imbalance, dealer hedging or liquidity provision. Order flow is the stream of buy and sell orders arriving at the market. The economic magnitudes are usually small per trade and, for retail traders, dominated by transaction costs. Institutions capture them as part of market making, the business of the dealers whose hedging Parts VI to IX study.

Microstructure edges live at the level of the order book (Chapter 2C):

Over a horizon $\Delta t$ of seconds to minutes, $f$ stands for whatever relation the data support, since the point of the formula is its inputs. The edge per unit is tiny, basis points or less, and viable only at scale with low transaction costs. An edge of 2 basis points on a $20,000 trade is $4, so any round-trip cost above $4 turns it negative.

What is in this category

• Short-term reversal over 1 to 5 days: liquidity provision after sharp moves.
• Dealer gamma effects: hedging flows around major strikes and around OPEX (Part VIII).
• Intraday momentum and mean reversion: patterns in order flow within the day.
• Pin risk and OPEX dynamics: options-related flows around expiry (Chapter 5F).
• Market-making spreads: quoting prices and providing immediacy.

Why it matters: not retail-tradable

Microstructure edges are essentially institutional-only. The edge per trade is usually 1 to 5 bps, which retail commission and slippage costs alone exceed. The infrastructure required is institutional too: low-latency execution, order-book data, and a large account for cross-margining, in which margin is set across related positions together.

A retail trader trying to capture microstructure edges is playing on a field where every other participant has structural cost advantages. Each of them is taking your money on each trade.

What it costs

For retail, the cost is just the transaction friction. For institutions, it is the rare adverse selection event when the flow pattern reverses: the "buyer" you were front-running was actually informed, and the news drops.

Where this book's chapters disagree

That caution is not shared by every chapter of this book. Parts VI to IX, which build the dealer's view of the market, give dealer positioning more weight. Their reason is the hedging mechanism they derive: a dealer who is long gamma sells into rallies and buys into falls to stay hedged, which damps moves. The market dashboard of Chapter 11A therefore reads the largest call strike above spot as potential resistance and the largest put strike below as potential support. The conviction score of Chapter 13B asks whether dealer positioning supports a trade's path.

This chapter's reasons for caution are the three above: vendor models disagree, 0DTE changed who holds what, and the measured hedging flow is small. On one point both views agree: scalping level by level is not supported by the evidence (Chapter 13F). The sign of dealer gamma is regime context, not a trade by itself.

6 · Why the taxonomy matters in practice

The source of an edge determines its persistence, its capacity, and how to use it:

A sleeve is the part of a portfolio given to one strategy. A factor tilt leans a stock portfolio toward shares with a chosen trait, such as value or quality. A tactical overlay adjusts the portfolio's existing exposure for a limited time rather than running as a strategy of its own. When you see a new edge, classify it first. This single classification step eliminates 80% of the noise in retail trading content.

7 · Mixed-source edges

Some edges combine several sources, which is fine as long as you can decompose them:

The decomposition matters because the parts age differently. In trend following, the behavioral part is at risk of decay and has already partly decayed, while the risk-premium and institutional parts are more durable. A purely behavioral reading of trend would be more pessimistic about its forward returns than the academic mixed view.

Most of the VRP is real risk premium, which does not decay. Some of it is institutional and might compress as hedging product structures evolve. A small behavioral part might decay further. Most of momentum is behavioral and has decayed substantially since publication. Its real crash-risk component explains why momentum can lose 73% in a quarter.

8 · From four sources to five categories

Chapter 10A introduces a five-category taxonomy: foundation premia, factor tilts, structural overlays, regime filters, and rejected. The sources map onto the categories roughly like this:

The mapping is not deterministic: VRP is a risk premium and sometimes implemented as a tilt, and BAB is institutional and often classified as a factor. Even so, the rough mapping helps you place a new edge quickly.

The key cognitive rule is this: before you can know how to size a strategy and how to kill it, you have to know what kind of edge it is. That is why this chapter comes before any single edge is examined in Part X. The reasoning runs from distributions (Chapter 1B) through the sources of edge in this chapter. It continues with why the Sharpe ratio misleads on negatively skewed strategies (Chapter 1D) and with realistic retail expectations, which differ by source (Chapter 13A). From there it runs through the conviction framework (Chapter 13B), the five categories (Chapter 10A) and the dossiers of Part X to the operational chapters of Part XIII.

9 · Check yourself