Signal Discovery Frameworks
Separating Informative Signals from Noise in Complex and Uncertain Markets
Financial markets generate an extraordinary quantity of information. Prices change continuously, companies publish financial statements, macroeconomic data arrives at regular intervals, investors communicate through transactions, and technological systems produce increasingly detailed observations of economic activity. The challenge facing the modern investor is therefore rarely a simple shortage of information. The more difficult problem is determining which information contains a signal.
A signal is an observable feature of the market or underlying economy that provides information about a future outcome of interest. It might relate to valuation, momentum, liquidity, profitability, investor behaviour, macroeconomic conditions, market structure, or an interaction between several variables. The existence of a potentially informative relationship, however, does not mean that it can be converted into a reliable investment strategy.
Markets contain enormous quantities of noise. Random variation can produce apparent relationships, historical coincidences can resemble persistent patterns, and sufficiently flexible models can discover structures that exist only within the data from which they were trained. Signal discovery is therefore not simply the search for variables that correlate with future returns. It is a framework for determining whether an observed relationship contains economically meaningful, statistically defensible, and potentially persistent information.
This distinction is fundamental because the objective is not to discover the largest historical pattern, but to identify information that has a reasonable prospect of surviving outside the sample in which it was discovered.
What Is an Investment Signal?
At its simplest, a signal is information that changes the probability distribution of a future outcome. An investor might observe that companies trading at unusually low valuations have historically generated different subsequent returns from companies trading at unusually high valuations. A quantitative researcher might identify a relationship between changes in earnings expectations and subsequent price behaviour, while a macro investor might observe that particular combinations of inflation, liquidity, and economic growth have historically corresponded with different asset-class outcomes.
In each case, the signal does not necessarily determine what will happen; rather, it changes the odds attached to different possibilities. This probabilistic interpretation is important because financial signals are rarely deterministic. Even a highly informative signal can be wrong on an individual observation, meaning that a strategy based upon it may experience losing trades, periods of underperformance, and false positives.
A useful signal should therefore be evaluated according to the information it adds to the decision-making process rather than according to whether it predicts every individual outcome correctly. The relevant question is whether knowledge of the signal improves the distribution of decisions and outcomes over a sufficiently meaningful sample.
Signal and Noise
The central challenge of signal discovery is distinguishing structure from randomness. Financial data contains both, and they are rarely labelled conveniently for the researcher. A company's improving profitability may contain meaningful information about its future economics, while daily price movements may simultaneously contain substantial stochastic variation unrelated to any persistent fundamental change.
The researcher observes the combined output and must infer whether an apparent relationship reflects a genuine mechanism or merely the statistical characteristics of a particular sample. This becomes increasingly difficult as the number of variables under consideration expands. If thousands of indicators are tested against thousands of securities across multiple time horizons, some relationships will inevitably appear statistically impressive purely through chance.
The existence of a pattern is consequently only the beginning of signal discovery. The more important question is whether the pattern survives attempts to disprove it and whether there is a convincing reason to expect the relationship beyond the data in which it was initially observed.
The Signal Discovery Process
A rigorous signal discovery framework can be understood as a sequence of increasingly demanding questions. The process begins with an observation or hypothesis that might emerge from economic theory, market structure, behavioural finance, previous empirical research, or exploratory analysis. The researcher then defines the variable precisely and establishes what outcome it is intended to explain or forecast.
Historical data can subsequently be examined to determine whether the relationship appears to exist, but the analysis must account for timing, data quality, survivorship, transaction costs, and other sources of distortion. If the relationship appears promising, it should be subjected to robustness testing across different samples, periods, assets, specifications, and market environments before being considered for practical implementation. Finally, the signal must be evaluated within the context in which it would actually be used. A statistically significant relationship that cannot be traded after costs may have little practical value, while a predictive variable that works only during one historical regime may be useful conditionally without constituting a reliable general signal.
Signal discovery is therefore best understood as a process of progressive scepticism. Each stage should make it increasingly difficult for an attractive but spurious relationship to survive.
Theory Before Data
One of the strongest foundations for signal discovery is economic or behavioural reasoning. A researcher who begins with a plausible mechanism has an important advantage over one who searches blindly through data for correlations.
Suppose a researcher hypothesises that investors systematically underreact to changes in corporate earnings expectations. That hypothesis provides a mechanism that can be tested. If the observed relationship subsequently survives rigorous analysis, the researcher has more than a statistical association; there is a possible explanation for why the relationship exists. This does not prove that the signal is genuine, but it provides a foundation for further investigation. Theory also helps constrain the search space. Without some prior reasoning, an enormous number of relationships can be tested, and the larger the search space becomes, the greater the probability of discovering accidental patterns.
Economic intuition therefore acts as a form of regularisation. It does not eliminate statistical uncertainty, but it can reduce the temptation to treat every historical correlation as meaningful.
Exploratory Discovery
Theory should not, however, become a constraint that prevents researchers from discovering relationships they did not anticipate. Exploratory analysis has an important role in quantitative research because markets are complex and not every useful relationship can be derived in advance from established theory.
Data can reveal structures that challenge existing assumptions, expose interactions that were previously overlooked, or suggest mechanisms that warrant further investigation. The difficulty is that exploratory findings carry a higher risk of false discovery because the researcher is effectively searching across a large hypothesis space. An exploratory discovery should therefore be treated as a hypothesis-generating result rather than immediate evidence of a deployable signal. The distinction between discovery and validation is essential. Discovery asks what might be happening, whereas validation asks whether the apparent relationship survives serious attempts to demonstrate that it is not real.
Confusing these two stages is one of the most common weaknesses in quantitative research.
Feature Engineering
Modern signal discovery often involves transforming raw observations into variables that better represent economically meaningful information. This process is commonly described as feature engineering.
A raw price series might be transformed into returns, volatility measures, momentum indicators, drawdown characteristics, or relative-strength measures. Financial statements can be converted into profitability, leverage, growth, valuation, or quality measures, while market data can be transformed into liquidity, dispersion, correlation, or flow-related variables.
Feature engineering is more than mathematical manipulation because the transformation determines what information the model is actually able to observe. A carefully designed feature can capture a relationship that would remain hidden in raw data, while an excessively complex feature may introduce noise or overfit historical observations. The researcher therefore needs to understand not only whether a feature improves historical performance but what information it represents and why that information should plausibly matter.
Cross-Sectional and Time-Series Signals
Signal discovery can operate across securities or through time. Cross-sectional signals attempt to distinguish between assets at a particular point in time, with valuation, profitability, quality, momentum, and relative strength providing common examples of variables that can be used to rank securities against one another.
Time-series signals instead focus on the behaviour of an individual asset or market through time. Trend, volatility, mean reversion, regime changes, and changes in liquidity can fall into this category. The distinction matters because the statistical structure of the problem differs between the two approaches. Cross-sectional signals depend upon relative comparisons and require careful treatment of sector, geography, size, and other structural differences, while time-series signals depend more heavily upon temporal dependence, regime stability, and the appropriate definition of historical information.
A comprehensive signal framework can examine both dimensions while recognising that evidence supporting one does not automatically validate the other.
Multi-Factor Signals
Individual signals are often imperfect. A valuation measure may identify inexpensive companies without distinguishing between temporary mispricing and structural deterioration, while momentum may capture persistent price trends without establishing whether those trends are sustainable. Quality measures can identify robust businesses but may fail to account for excessive valuations.
Combining signals can potentially produce a more informative representation of an investment opportunity, but the objective should not simply be to add as many variables as possible. If multiple signals measure essentially the same underlying phenomenon, their apparent diversification may be misleading. Conversely, signals derived from genuinely different mechanisms may provide complementary information.
The important question is therefore whether each component contributes incremental information. A strong multi-factor framework is not necessarily the one with the greatest number of factors, but rather the one in which the combination improves inference without introducing unnecessary complexity or redundant exposures.
Nonlinear Relationships
Many traditional investment models implicitly assume that relationships are linear, yet real-world financial relationships are frequently more complicated. A moderate increase in leverage may have relatively little effect while a company's balance sheet remains resilient, whereas additional leverage beyond a particular threshold can dramatically increase financial fragility.
The same principle applies to market variables. The relationship between liquidity and volatility may change when markets become stressed, while a variable that appears weakly related to returns under normal conditions can become considerably more informative during periods of market dislocation. Signal discovery frameworks should therefore allow for nonlinear relationships where economic reasoning supports them. The difficulty is that additional flexibility also increases the risk of overfitting because a highly flexible model can reproduce historical data extremely well while possessing little ability to generalise beyond it.
The value of complexity consequently depends upon whether the additional structure corresponds to something meaningful in the underlying system.
Interactions Between Signals
The information contained in one variable can depend upon another. Momentum may behave differently in high-volatility and low-volatility environments, valuation may have different implications when liquidity is abundant compared with when financing conditions are tightening, and an earnings revision may be more informative when investor positioning is unusually concentrated.
These interactions can contain valuable information because financial markets are rarely governed by independent variables. The challenge is distinguishing meaningful interactions from accidental combinations. A useful interaction should therefore have a plausible economic mechanism and demonstrate some degree of stability across different samples and environments. Otherwise, the additional complexity may simply give the model another opportunity to fit historical noise.
Regime-Dependent Signals
One of the most important considerations in signal discovery is that a signal's effectiveness can depend upon the state of the broader system. A relationship that works under one market regime may weaken or reverse under another as interest rates, inflation, liquidity, volatility, economic growth, investor positioning, and market structure change.
This means that the question "Does this signal work?" can be insufficient. A more informative question may be "Under what conditions does this signal contain information?"
Regime-aware signal discovery attempts to answer that question by considering whether observations can be meaningfully differentiated according to the state of the broader market and whether the signal behaves differently within those states. This approach is particularly relevant to financial markets because the underlying generating process can evolve through time. A signal that appears weak across an entire sample may be highly informative within a particular regime, while a signal that appears strong on average may owe most of its historical performance to a small number of specific environments.
Statistical Significance Is Not Economic Significance
A statistically significant result is not necessarily an economically useful result. A sufficiently large dataset can make a very small effect appear statistically significant, while a potentially important relationship may appear statistically weak because the sample is small or unusually noisy.
Investors therefore need to distinguish between statistical evidence and practical relevance. A signal may have a measurable relationship with future returns while offering little economic value after transaction costs, taxes, market impact, financing costs, and other implementation constraints. The ultimate question is whether the signal improves the decision in an economically meaningful way. Answering that question requires moving beyond statistical inference toward portfolio construction, risk management, and implementation analysis.
Multiple Testing and the False Discovery Problem
Signal discovery becomes particularly vulnerable to false discoveries when researchers test large numbers of hypotheses. If thousands of indicators are evaluated and only those producing attractive historical results are retained, some will inevitably appear successful simply because of random variation.
This is the multiple-testing problem; the more opportunities a researcher gives randomness to produce an impressive result, the greater the number of impressive results randomness can produce. Large-scale research is not inherently flawed, but the evidential standard must rise as the search space expands.
Correction methods, holdout samples, false-discovery controls, economic priors, robustness testing, and independent validation can all help reduce the risk. More fundamentally, researchers should maintain an explicit distinction between the number of hypotheses investigated and the number of relationships that ultimately appear successful. The path through which a signal was discovered is itself relevant evidence.
Backtesting and the Problem of Overfitting
Backtesting provides an essential tool for evaluating historical signal behaviour, but it can also create false confidence. A model that fits historical data extremely well may simply have learned the peculiarities of the dataset rather than the underlying structure that generated it.
Overfitting occurs when a model captures noise or idiosyncratic historical relationships rather than a mechanism that can generalise beyond the development sample. The danger increases with model flexibility because every additional parameter, transformation, threshold, interaction, and optimisation decision creates another opportunity for the model to adapt to historical noise. A robust signal should therefore be subjected to out-of-sample testing. The model should be evaluated on observations that were not used during its development and, where possible, across different periods, market environments, geographic regions, asset classes, and implementation assumptions.
The purpose is not to prove that a signal will work indefinitely, because no historical test can provide such a guarantee. Instead, the objective is to determine whether the available evidence is consistent with a relationship that has a reasonable prospect of generalising.
Data Snooping and Researcher Degrees of Freedom
A related problem arises from researcher degrees of freedom. During development, a researcher may make many seemingly reasonable decisions concerning which securities to include, which period to analyse, how to define the signal, which outliers to remove, which threshold to use, and which performance metric to optimise.
Each choice can influence the eventual result.
When enough decisions are made after observing the data, the final strategy may become highly tailored to the historical sample. The resulting performance can therefore reflect the researcher's accumulated choices rather than a genuinely persistent relationship in the market.
This is why reproducibility and research discipline matter. A credible signal discovery framework should preserve an audit trail showing how a signal was conceived, modified, tested, rejected, and ultimately selected. As such, the research process itself becomes part of the evidence.
Economic Mechanism and Causal Reasoning
Correlation can be useful for discovering signals, but causal reasoning provides a stronger foundation for understanding them. Suppose a variable predicts future returns. There are several possible explanations:
it might capture a genuine economic mechanism
proxy for an omitted risk factor, represent investor behaviour
reflect market microstructure
or simply correlate with another variable that actually drives the outcome
Understanding the mechanism matters because relationships without mechanisms can be fragile. If a signal exists because investors systematically behave in a particular way, changes in technology, regulation, or market participation may alter it. If it exists because the signal compensates investors for bearing a particular risk, the relationship may persist as long as the underlying risk remains priced.
Causal reasoning therefore helps investors understand not only whether a signal exists, but why it might exist and under which circumstances its explanatory power should remain intact.
Signal Decay
No signal should be assumed to remain constant indefinitely. Once a signal becomes known, investors may attempt to exploit it, causing capital to flow toward the opportunity and competition to increase. The resulting pressure can reduce the available return even when the original relationship remains economically meaningful.
Signal decay can also occur because the underlying market structure changes. Technology can reduce execution costs, regulation can alter incentives, new information sources can make previously scarce information widely available, and institutional participation can change liquidity and price formation.
A signal that worked historically may therefore weaken without disappearing entirely. This creates an important distinction between signal existence and signal strength. The relationship may become smaller, slower, more conditional, or more expensive to exploit rather than simply ceasing to exist.
Capacity and Implementation
A signal can possess predictive power without being scalable. An opportunity may work for a relatively small amount of capital but deteriorate as more capital attempts to exploit it because market impact increases, liquidity declines, and execution costs consume the expected return.
Capacity is therefore an important component of signal evaluation; and theoretical alpha and investable alpha are not necessarily equivalent. A signal discovery framework that ignores implementation can systematically overestimate the value of a strategy because it treats a theoretical relationship as though it could be captured without friction.
Institutional research must instead consider the size, liquidity, turnover, trading costs, financing requirements, and market impact associated with converting the signal into actual positions.
Robustness as a Form of Evidence
Robustness is one of the most important concepts in signal discovery because a relationship that works only under one narrowly defined specification should be treated differently from one that survives reasonable changes in methodology.
Researchers can examine whether a relationship persists when variables are defined slightly differently, when extreme observations are removed, when different periods are analysed, or when alternative reasonable portfolio-construction methods are employed.
The objective is not to demand identical results under every possible specification, since financial relationships are rarely that stable. Instead, robustness asks whether the central conclusion remains reasonably intact when reasonable assumptions change. As, a signal that survives these tests provides stronger evidence than one whose entire performance depends upon a single parameter choice.
Machine Learning and Signal Discovery
Machine learning has expanded the possibilities of signal discovery by allowing researchers to model nonlinear relationships, interactions, high-dimensional datasets, and complex structures that traditional models may struggle to represent.
This creates significant opportunities, but it also introduces significant risks. The flexibility of modern models means they can discover extremely complex patterns within historical data, and without careful validation the model may learn the dataset rather than the market.
Interpretability can also matter; a highly predictive model whose behaviour cannot be understood may be difficult to distinguish from a sophisticated form of overfitting, particularly when the model has been selected from a large population of competing alternatives. Machine learning should therefore be regarded as a tool for inference rather than an automatic source of investment insight. The quality of the research design remains more important than the sophistication of the algorithm.
Signal Ensembles and Model Diversity
One response to the uncertainty surrounding individual signals is to combine multiple signals or models. The underlying logic resembles diversification: if several signals capture different mechanisms and are imperfectly correlated, their combination may produce a more stable information set than reliance upon any single relationship.
Model diversity can be particularly valuable when different models respond differently to changing market environments. A momentum model, valuation model, liquidity model, and behavioural model may each contain partial information, and their combination can potentially provide a broader representation of market conditions. Diversification of models should not, however, be confused with diversity of information. Five signals derived from essentially the same underlying price behaviour do not necessarily provide five independent sources of evidence.
Thus, the objective should be diversity of mechanisms rather than simply diversity of formulas.
Signal Discovery as an Ongoing Process
Signal discovery should not end when a strategy is deployed because markets continue to evolve and the relationship between a signal and its target variable can change. A mature research framework should therefore monitor signal performance, stability, implementation costs, exposures, crowding, and environmental dependence through time.
This creates a feedback loop between research and deployment. Live performance generates new evidence, that evidence informs further research, and research can identify potential changes in the underlying mechanism that influence subsequent implementation. The process is continuous, but continuous research does not mean continuous strategy modification. A framework must distinguish between genuine evidence that the underlying mechanism has changed and temporary performance variation that remains consistent with the original thesis.
This distinction is essential because an investment process that adapts too aggressively can become a mechanism for chasing noise rather than learning from information.
The MorMag Perspective
At MorMag, signal discovery is best understood as a disciplined process of separating potentially informative structure from the immense quantity of noise generated by financial markets.
The objective is not to maximise the number of signals discovered or to identify the model that produces the most impressive historical backtest. Instead, the aim is to determine whether an observed relationship contains information that is economically meaningful, statistically defensible, robust across reasonable specifications, and potentially useful within an actual investment process.
This requires a clear distinction between discovery and validation. Discovery is exploratory and permits data to reveal relationships that may not have been anticipated in advance. Validation is deliberately more adversarial because it asks whether those relationships survive attempts to demonstrate that they are products of chance, selection effects, overfitting, unstable regimes, implementation costs, or hidden common exposures. This distinction is particularly important in complex financial systems because a sufficiently flexible research process can almost always find something interesting. The difficult part is determining whether that something matters.
MorMag's broader research philosophy therefore treats signals as probabilistic evidence rather than deterministic instructions. A signal should alter the assessment of an investment opportunity without eliminating the uncertainty surrounding it. Its meaning depends upon the context in which it appears, including market regime, liquidity conditions, investor positioning, valuation, volatility, and the behaviour of competing strategies. A signal is therefore not simply a number extracted from a dataset. It is information embedded within a system. This perspective has important implications for quantitative research infrastructure. Signal discovery should incorporate statistical testing alongside regime analysis, robustness testing, portfolio awareness, behavioural interpretation, risk assessment, and consideration of how a discovered relationship might change once other investors begin exploiting it.
The ultimate objective is not to build a machine that predicts markets with certainty; it is to develop a decision framework capable of extracting useful information from uncertainty while remaining conscious of the possibility that any particular relationship can weaken, disappear, or prove illusory. The strongest signal discovery framework is therefore not necessarily the one that finds the greatest number of patterns. It is the one that is most effective at determining which patterns deserve to be believed, which deserve further investigation, and which should be rejected.
Conclusion
Signal discovery lies at the heart of modern investment research because markets contain vastly more information than any investor can process directly. The challenge is therefore not simply finding information, but determining which information changes the probability of future outcomes in a meaningful and persistent way.
A rigorous framework must navigate the distinction between signal and noise, theory and exploration, statistical significance and economic significance, discovery and validation, and historical performance and future relevance. It must also account for overfitting, multiple testing, researcher degrees of freedom, implementation costs, capacity, regime dependence, and the possibility that successful signals attract competition and subsequently decay. These challenges become even more important as quantitative tools become increasingly powerful. The ability to search larger datasets and construct more flexible models expands the possibility of discovering genuine structure, but it also expands the opportunity for randomness to masquerade as information.
The fundamental problem is therefore epistemic:
how do we know that a signal is real?
There is rarely a definitive answer. Instead, confidence must be built through converging evidence, including a plausible mechanism, appropriate statistical analysis, robustness across samples, sensible implementation assumptions, independent validation, and continued monitoring through changing environments.
Markets are not static datasets waiting to reveal their secrets; they are adaptive systems populated by participants who learn, compete, imitate, and respond to the very signals researchers are attempting to discover. A relationship that appears powerful in one environment may consequently weaken when the surrounding system changes; the most valuable signal may therefore be less about predicting a particular price movement than about improving the quality of the decision-making process itself.
Signal discovery, at its deepest level, is an exercise in disciplined inference. It is the attempt to extract structure from complexity without mistaking complexity for structure, to learn from data without becoming captive to it, and to identify information that remains useful after the market, the model, and the researcher have all been given opportunities to be wrong.
