For 1,400 years, Ptolemy's geocentric model of the solar system served astronomy reasonably well. It predicted planetary positions accurately enough for navigation and agriculture. When observations drifted from predictions, astronomers added epicycles, smaller circular motions superimposed on larger ones, and accuracy returned. The model could absorb any anomaly by adding complexity.
This is the signature of descriptive models: they match patterns without explaining mechanisms. Copernicus didn't offer a more accurate model so much as a model that explained why planets moved as they did. The shift from geocentrism to heliocentrism wasn't primarily about better predictions. It was about the difference between curve-fitting and causal understanding.
Factor investing in finance has the structure of Ptolemaic astronomy. When the Capital Asset Pricing Model—CAPM, the foundational theory that an asset's expected return is proportional to its sensitivity to the overall market—couldn't explain returns, Fama and French added factors. When three factors weren't enough, they added two more. Smart beta proliferated into dozens of variations. Each new factor responded to the last model's failures: momentum because value didn't capture everything, quality because momentum crashed, low volatility because the whole framework seemed inverted.
These models can match historical returns with reasonable accuracy. But matching isn't explaining. Factor investing—the practice of tilting portfolios toward characteristics like value, size, or momentum that have historically earned excess returns—has the same structure as Ptolemaic astronomy: it describes patterns without identifying the mechanisms that produce them. I think the entire framework is building in the wrong direction, adding epicycles to preserve a structure that mistakes descriptions for causes. And when you're pattern-matching descriptions rather than identifying causal mechanisms, you're always one regime change away from confusion.
Figure 1 Epicycles to Ellipses: Same Motion, Different Explanations
Ptolemaic Model
Heliocentric Model
Both models produce the same apparent planetary motion as seen from Earth. The Ptolemaic model requires circles upon circles (epicycles) to match observations. The heliocentric model explains the same motion with simple elliptical orbits—the apparent complexity emerges from our moving vantage point. Factor models in finance have the Ptolemaic structure: they describe patterns without explaining why those patterns exist.
The Platonic Trap
Columbia Business School, my alma mater, has a heritage of value investing woven through its curriculum. Benjamin Graham taught there and Warren Buffett was his student. During my time as an MBA student starting in 2017, I watched the stock market's gains accrue almost entirely to companies that value frameworks would have screened out. I sometimes wondered how my professors' portfolios were performing as they taught these classes. Today there's no ambiguity: anyone practicing what they preached would have captured barely half the market's gains since I started school. The value premium didn't just compress. It inverted.
Value investing and the 60/40 portfolio share an underlying assumption: that certain relationships in markets are stable enough to treat as permanent features. Cheap stocks outperform expensive ones. Bonds diversify equity risk. These patterns held for decades, which made them easy to mistake for something deeper.
Plato described Forms as the true reality underlying appearances, the perfect abstractions that physical objects merely approximate. When we treat a pattern as a Form, we're often capturing an artifact of the conditions that produced it rather than anything stable enough to rely on.
The value premium didn't emerge from a fundamental law of markets. It emerged from an era of falling interest rates that favored duration, accounting conventions that created predictable mispricings, and systematic behavioral tendencies among investors who underweighted certain characteristics. When those conditions changed, the premium compressed. The pattern was real, but it was never a Form. It was a shadow cast by a particular configuration of market structure, and when the structure shifted, the shadow changed shape. Value stocks underperformed growth by over 30 percentage points in 2020 alone, the kind of drawdown that isn't supposed to happen to a "risk factor" that commands a premium.
Figure 2 The Value Premium Disappearance
Cumulative returns of Russell 1000 Growth vs. Value indexes since 1995, both starting at 100%. Through the dot-com bust, value actually outperformed. But since 2017, the lines have diverged dramatically. Growth returned approximately 2,600% cumulative by end of 2024; Value returned approximately 1,670%. Anyone following traditional value frameworks would have captured barely half the market's gains. The pattern that held for decades wasn't a Form—it was an artifact of conditions that changed.
The 60/40 portfolio had a similar problem. Stock and bond returns were negatively correlated for a long period, which made bonds an effective hedge for equity risk. But that correlation wasn't a law of nature. It was an artifact of a particular monetary regime, a specific range of inflation dynamics, and the positioning of policy rates relative to growth expectations. When the regime shifted, diversification benefits vanished. In 2022, stocks and bonds fell together, and the correlation flipped positive for the first time in two decades. Investors who had modeled the negative correlation as a stable feature of markets discovered it was conditional on a structure that no longer held.
I find this a useful frame for thinking about why sophisticated investors get caught in drawdowns they didn't anticipate. They modeled risk using relationships they believed were stable, and the relationships turned out to be conditional on a market structure that was itself changing.
Figure 3 Factor Proliferation: Adding Epicycles
Each time factor models failed to explain returns, the response was to add more factors. The rings below each event represent the number of "epicycles" in the model. By the 2020s, smart beta had proliferated into dozens of variations. More epicycles don't become heliocentrism—they just become a more elaborate way of describing patterns without explaining them.
What Actually Moves Prices
If factors are epicycles to CAPM's geocentrism, what's the heliocentric alternative? I've spent some time thinking about the root causes of pricing factors, and it will take a few articles to fully lay out. But I'll start with the four categories of causal mechanisms that I think actually drive asset prices.
Three of these mechanisms are persistent, each causally distinct. The fourth is an emergent phenomenon that appears under specific conditions.
Start with a simple causal model: something changes in the world, which affects a company's fundamentals, which updates investor beliefs about value, which moves price. Most of finance implicitly assumes this chain. But examining it closely reveals that different mechanisms enter the chain at different points, and some bypass it entirely. Mechanisms that flow through the full chain pull on price by changing what investors think an asset is worth. Mechanisms that skip straight to price push it through mechanical flows, moving price without changing anyone's beliefs about value. The diagrams below make this visible: pull factors trace the complete path from cause to fundamentals to beliefs to price; push factors arc directly to price, leaving fundamentals and beliefs grayed out. This isn't a cosmetic difference. It determines whether a price movement carries information about value or simply reflects market plumbing.
Endogenous factors are firm-specific pull dynamics. Management decisions, operational execution, product-market fit, capital allocation—these change fundamentals, which updates beliefs, which moves price. Fundamental analysis tries to identify these, and they're real, but they largely diversify away in portfolios. Own thirty stocks and firm-specific variance starts canceling out.
Exogenous factors are structural pull dynamics outside any individual firm's control. Macroeconomic conditions, sector dynamics, regulatory shifts, technological displacement, demographic trends—these also flow through fundamentals and beliefs to reach price. Unlike endogenous factors, they don't diversify away. A portfolio of retail stocks remains exposed to consumer spending regardless of how many retailers you own. A portfolio of banks remains exposed to the yield curve. The exposure isn't a choice. It's embedded in what these businesses are.
Push factors are mechanical flows that move prices independent of any view on fundamentals—the curved arrow in the diagram that bypasses beliefs entirely. Index inclusion forces buying regardless of valuation. Central bank balance sheet operations reprice duration assets through portfolio rebalancing. ETF creations and redemptions create predictable patterns of buying and selling pressure. These aren't information about value. They're structural features of market plumbing, and they can move price without anyone changing their mind about what the asset is worth.
The three mechanisms interact but remain causally distinct. Consider Nvidia in 2023. The stock rose roughly 240% over the year. Some portion of that reflected genuine endogenous improvement: the company's positioning in AI infrastructure, execution on product roadmap, expanding margins. Some portion reflected exogenous forces: the broader AI investment cycle, sector rotation into technology, falling rate expectations in the back half of the year. And some portion was pure push: Nvidia's rising weight in the S&P 500 forced index funds to buy more shares mechanically, and its inclusion in various AI-themed ETFs created additional flow. All three mechanisms contributed to the same price movement, but they have different implications for persistence. The endogenous gains might compound if execution continues. The exogenous gains depend on whether the AI cycle sustains. The push gains could reverse if index weights rebalance or thematic flows shift.
Figure 4 Variance Decomposition: Nvidia 2023
The same price movement can have completely different causal compositions. Nvidia's 2023 gains were primarily endogenous (execution) and exogenous (AI cycle). A utility like Duke Energy is dominated by exogenous factors (rates). GameStop in 2021 was mostly reflexive—the market was trading disagreement itself, not fundamentals. Understanding which mechanism drives returns matters because persistence differs: endogenous gains can compound, exogenous gains depend on conditions, push gains often reverse, and reflexive gains collapse when uncertainty resolves.
Understanding which mechanism is driving a price movement matters because the appropriate response differs. Endogenous movements might persist if they reflect genuine improvement in the business. Exogenous movements depend on whether macro conditions continue. Push movements often reverse once the mechanical flow completes.
When Disagreement Becomes the Trade
There's a fourth phenomenon, but calling it a fourth factor would be a category error. It's better understood as a variable portion of variance that exists in every market but sometimes dominates.
Reflexive dynamics appear when price itself becomes the signal traders respond to, rather than external information about fundamentals. In normal markets, the causal arrow points one direction: information about the world updates beliefs, and beliefs move price. In reflexive markets, a feedback loop emerges: price moves update beliefs about where price is heading, which moves price further. The loop is self-referential. Traders aren't disagreeing about what the company is worth—they're disagreeing about what other traders will do.
This isn't a binary. Every market has some reflexive component—the question is how much of the variance it explains. A stable utility stock might be 5% reflexive; price movements are mostly information about rates and regulation flowing through beliefs. A meme stock at peak frenzy might be 80% reflexive; almost no one is trading a view on fundamentals. The reflexive portion waxes and wanes as conditions change.
Several signals indicate when reflexive dynamics are growing. Price-volume feedback is one: in normal markets, volume spikes on news; in reflexive markets, volume spikes on price moves themselves, with no news required. Return autocorrelation is another: reflexive regimes show momentum clustering, where positive returns predict positive returns at short horizons. Options markets reveal it too: when implied volatility for short-dated options diverges dramatically from longer-dated, people are trading expected price moves rather than fundamental uncertainty. And when price movements decouple from news sentiment—big moves on no news, or counterintuitive reactions to earnings—the market has likely disconnected from fundamentals.
Figure 5 Detecting Reflexive Dynamics
No single metric cleanly separates reflexive from fundamental dynamics. But when multiple signals align—volume trading on price rather than news, momentum clustering, options term structure inversion, decoupling from fundamental catalysts—the reflexive portion of variance is likely elevated. These signals don't tell you where price should be. They tell you that the market has temporarily stopped asking that question.
Reflexive dynamics alone don't create bubbles. A stock can trade reflexively for months—high volatility, momentum clustering, disconnection from fundamentals—without prices running away. GameStop in early 2021 was intensely reflexive, but it was a spike, not a bubble. The distinction matters.
Bubbles emerge when reflexive dynamics combine with additional conditions that create runaway feedback. Four conditions seem necessary:
First, asymmetric constraints. Long positions are typically unconstrained—anyone can buy. Short positions face friction: borrow costs, margin requirements, career risk for professionals. This asymmetry creates a ratchet. Reflexive upward pressure isn't balanced by reflexive downward pressure. Skeptics get squeezed out rather than rewarded.
Second, participant inflow. Reflexive dynamics in a fixed pool of traders are self-limiting—the same dollars just slosh around. But when rising prices attract new participants—retail investors seeing headlines, funds chasing performance, tourists from adjacent markets—fresh capital enters the loop. The participant pool expansion is itself driven by price, creating a second-order feedback.
Third, narrative elasticity. Fundamentals provide an anchor, but only if the narrative is constrained. When the story becomes elastic—"it's not about current earnings, it's about TAM a decade from now"—the anchor loosens. Price can run further before hitting any reality constraint that forces a reassessment. Elastic narratives buy time for reflexive dynamics to compound.
Fourth, leverage and forced flows. When rising prices trigger mechanical buying—margin calls on shorts, delta hedging by options dealers, index inclusion—push dynamics layer on top of reflexive dynamics. The combination is explosive. Price moves that would be bounded by beliefs alone become self-amplifying through structure.
When all four conditions align with elevated reflexivity, you get a bubble. Remove any one, and you get something less: a spike (GameStop—no elastic narrative), a momentum run (AI stocks 2023—constraints less binding), or a slow grind (value traps—no participant inflow). The bubble is an emergent property of the configuration, not a separate phenomenon.
Bubbles end when one of the conditions breaks. Constraints become symmetric—shorting gets easier, or a liquid derivative market opens. Participant inflow stops or reverses—retail exhaustion, fund redemptions. The narrative gets reality-checked—an earnings miss, fraud revelation, or macro shock. Leverage unwinds—margin calls cascade. Often several break simultaneously, which is why collapses are faster than buildups. The conditions are fragile and correlated.
Recognizing this structure matters for portfolio management. When you see elevated reflexivity, the question isn't "is this a bubble?" but "which conditions are present?" If all four are aligned, position sizing and hedging need to reflect that the distribution has fat tails in both directions. If only one or two are present, the reflexive dynamics are more likely to burn out than to run away.
The Properties That Aren't Properties
This framework clarifies something important about two concepts that traditional finance treats as fundamental: beta and the equity risk premium.
Beta, in standard models, is presented as an intrinsic property of assets, a measure of systematic risk that commands compensation. But beta isn't intrinsic to anything. It's an artifact of how you decompose variance.
A stock's total price variance comes from endogenous, exogenous, push, and potentially reflexive sources. Beta measures a ratio: how much of the variance traces to exogenous factors relative to the market's exogenous variance. A high-beta stock isn't riskier in some fundamental sense. It's a stock where macro and sector forces explain a larger share of price movement than firm-specific forces do.
Figure 6 Exogenous Exposure by Asset
Each asset's variance can be decomposed into global exogenous factors (the coral band) and everything else. A Treasury ETF is almost pure global exogenous—its price moves with rates and macro conditions. The S&P 500, as the market itself, is similarly dominated by global factors. Duke Energy has moderate global exposure through rate sensitivity, but also faces regulatory and operational variance. Nvidia's variance is mostly not global exogenous—AI sector dynamics, firm execution, and reflexive disagreement about AI's trajectory dominate, even though these feel like "macro" forces. A pre-revenue biotech is barely touched by global macro; trial outcomes and sector sentiment drive nearly everything. The animation shows non-exogenous portions ebbing and flowing—the more an asset extends beyond the band, the more idiosyncratic its dynamics.
How do you actually measure these exposures? The decomposition isn't directly observable, but several identification strategies can tease apart the sources. Natural experiments help isolate individual mechanisms: Fed announcements are pure exogenous shocks, index inclusions are pure push, earnings surprises are mostly endogenous. Measuring an asset's sensitivity to each type of event reveals its exposure to that mechanism. Cross-sectional covariance structure offers another angle: movements that correlate across all assets are likely global exogenous, movements correlated within sectors but not across them are sector-level exogenous, and movements uncorrelated with peers are endogenous or reflexive. For reflexive dynamics specifically, disagreement metrics serve as detectors—analyst estimate dispersion, the ratio of implied to realized volatility, short interest combined with unusual volume. When these spike relative to fundamental news flow, the asset is likely in a reflexive regime. None of these methods is perfect, but combining them with careful attention to timing signatures and flow data can produce estimates precise enough to matter for portfolio construction and risk management.
This explains why beta isn't stable over time. As a company matures, endogenous variance often falls relative to exogenous exposure. A startup's price movements are dominated by questions about whether the product will work. A mature utility's price movements are dominated by where interest rates are heading. Beta rises not because something intrinsic changed but because the composition of variance sources shifted.
The equity risk premium has a similar structure. The standard story is that investors demand compensation for bearing systematic risk, and the premium is that compensation. I think this gets the causality backwards.
Markets don't price expected returns. They price distributions, with particular attention to the left tail. When investors evaluate outcomes using something like CVaR, pricing at a tail percentile rather than the mean, assets with right-skewed distributions systematically trade below their expected-value price. The gap between CVaR-based pricing and mean-based pricing shows up empirically as a "premium."
This isn't compensation for an abstract concept called risk. It's a mathematical consequence of how preferences interact with distribution shapes. The premium varies across assets not because of their beta but because of their return distributions. This becomes central to portfolio construction, which I'll address in a subsequent piece on the geometry of preference.
Volatility as Symptom
There's a broader confusion embedded in how finance treats volatility, and it connects to the Ptolemaic problem.
Volatility is a descriptive statistic. It summarizes the dispersion of realized returns. But it doesn't cause anything. The causes are the mechanisms that produce the returns volatility describes.
Treating volatility as a factor reverses the causality. Two assets with identical volatility can have completely different compositions. One might be volatile because of firm-specific uncertainty about a drug trial. Another might be volatile because of macro exposure to commodity prices. A third might be volatile because it's trading in a reflexive regime where disagreement has overwhelmed fundamentals. The number is the same, but the analysis and likely persistence differ entirely.
Factor models that include volatility as an input are adding epicycles. They're describing a pattern without asking what produces it. When the underlying mechanism shifts, the pattern breaks, and the model provides no explanation for why.
The Bayesian Inversion
There's a deeper methodological issue at stake, one that separates how finance typically thinks about models from how it should.
The standard approach treats data as primary and models as summaries. You observe returns, correlations, factor loadings. You fit a line, estimate parameters, decompose variance. The model is a description of the data—a compression of what happened into coefficients that might predict what will happen next.
This gets the epistemology backwards. The more productive framing asks: what process generated this data? Instead of fitting a line to points, ask what mechanism most likely drew these points. A fitted line is a description. A generative process is a theory. Descriptions can't extrapolate (or rather, they shouldn't, but often incorrectly do). Theories, however, can.
Figure 7 Upstream vs. Downstream
Causal forces (macro conditions, geopolitics, technology, rates) flow like tributaries upstream. They mix and interact midstream, creating sector effects and correlations. Downstream, we observe returns—the aggregate of everything. Traditional models measure the river's flow. Bayesian models ask what's happening at the source. The river is always a mixture of all tributaries; we're never in "one tributary's world."
The practical consequence: move upstream. Factor models sit downstream, summarizing the joint distribution of returns. Causal models sit upstream, specifying the forces that produce those returns. When the world changes, downstream descriptions break because they never understood what produced the patterns they described. Upstream models can adapt because they track the actual generators.
Complexity in the Right Direction
Models should be as complex as they need to be. This cuts both ways, and finance has managed to err in both directions simultaneously.
CAPM was too simple. One factor couldn't capture the distinct mechanisms that actually move prices. But the response wasn't to ask what those mechanisms were. It was to add factors until historical fit improved. The result is models that are complex in the wrong dimension: many parameters, no causal structure.
An overfitted model isn't just likely to fail out of sample. It becomes a representation of a point in time rather than a framework for understanding change. A model with seventeen factors calibrated to data from 1990 to 2010 is a photograph of that era's market structure. It offers no mechanism for anticipating what happens when the structure shifts.
The heliocentric move in astronomy wasn't about simplicity in a naive sense. Copernicus's model was initially more complex than Ptolemy's in certain respects. But it was complex in a productive direction: it described mechanisms that had causal force, which meant it could accommodate new observations without endless parameter additions. Kepler could refine it. Newton could explain it. The complexity was generative rather than defensive.
Finance needs something similar. Fewer factors, but grounded in actual mechanisms. Simpler descriptions of what we observe, but richer explanations of why we observe it. The goal isn't to predict every correlation. It's to understand market structure well enough that you're not blindsided when surface patterns shift.
The epicycles served astronomy for fourteen centuries. Factor models have served finance for perhaps five decades. In both cases, the framework worked until it didn't, and working was never the same as understanding.