The quantitative investing community continues to wrestle with a central tension: how much trust should be placed in backtesting versus in models that encode causal mechanisms? Authors such as Marcos López de Prado and Vincent Zoonekynd have urged practitioners to look beyond historical returns and ask why a strategy worked rather than assuming the past will repeat. That critique is valuable, but the choice is rarely binary. In real-world decision making—where information is incomplete and time is limited—both associational signals and structural explanations play roles.
Understanding the conditions under which each is useful is the practical problem.
To sharpen that judgment, it helps to think in layers. At the base is association: empirical patterns detected by tests and screens. Above that sits causal reasoning: hypotheses about mechanisms and transmission. Above both is reflexivity: the idea that market participants change behavior in response to signals, which then alters the phenomenon being predicted. Treating these as complementary rather than opposing approaches leads to clearer due diligence: when to accept a correlation as actionable, when to demand a structural model, and when to anticipate that using a signal will change its effectiveness.
Table of Contents:
Why historical fits are useful but limited
There are many reasons backtests remain central to quant research. They are quick to implement, often capture predictive associations that outperform naïve benchmarks, and help sort candidate signals at scale. In contexts where drivers are partially observable or resources to build mechanistic models are scarce, an association can be the most pragmatic guide. Yet association alone is an absence of explanation: a statistical relationship without a plausible transmission mechanism is fragile to sample selection, regime shifts, and data mining. Robust practice therefore layers association with tests for overfitting, out-of-sample validation, and sensitivity analysis, but those steps still do not substitute for thinking about mechanism.
A three-layer framework for clearer reasoning
Adopting a structured taxonomy clarifies what kind of evidence supports a trade. The first layer, association, answers: does this signal predict returns in the data? The second layer, causal, asks: what mechanism could create that pattern and how durable is it? The third layer, reflexive, considers: if many participants adopt this signal, how will behavior, flows, and liquidity evolve? Making these distinctions explicit helps investment teams allocate effort: invest in mechanistic modeling where durable structure exists, rely on associational rules where speed and observability favor them, and design position sizing and monitoring to mitigate reflexive crowding.
Epidemiology as a useful analogy
Think of epidemiology: analysts do not treat disease spread as a naked time series. Models such as SIR and SEIR encode compartments and transitions—susceptible, exposed, infectious, recovered—and then estimate parameters statistically. The prior causal structure guides interpretation and policy. Financial researchers can borrow this habit: when mechanisms like leverage amplification, funding constraints, inventory dynamics, or passive flow effects are credible, build them into a model rather than relying solely on historical correlation. That reduces the risk of mistaking transient comovement for a sustainable edge.
Accounting for reflexivity in markets
Markets differ because they are adaptive systems. Prices influence beliefs; beliefs influence prices. A compelling narrative can attract capital, and that capital can validate or reverse the narrative. This reflexivity implies that causation in markets often includes feedback loops between expectations, flows, and fundamentals. Dynamic models that represent stocks, flows, delays, and feedback are better suited to capture these processes than static regressions. Recognizing reflexivity helps explain why a relationship that held in the past may decay as it becomes widely exploited.
Practical steps for disciplined quantitative practice
Operationally, teams should follow a few simple rules. First, use causal knowledge where it exists and make the assumptions explicit. Second, prefer dynamic models when accumulation, delays, or feedback are central; these capture how today’s state shapes tomorrow’s outcomes. Third, when relying on associational signals, implement stringent out-of-sample tests, stress scenarios, and capacity constraints to gauge fragility. Finally, monitor adoption and liquidity metrics to detect reflexive crowding early and adjust exposures before transient correlations break down.
The goal is not to eliminate association-based methods or to fetishize causal models. Rather, the objective is transparency about which layer is driving investment decisions and how layers interact. Quantitative investing that combines fast statistical screens, mechanistic insight where available, and explicit treatment of reflexivity will be more resilient. That plural approach acknowledges the limitations of pure backtests while preserving the pragmatic strengths of data-driven discovery.
