What Is a Signal?

01The Core Idea

A signal is a piece of information that has some relationship with a future outcome you care about. That's it. Everything else is engineering.

In finance specifically, a signal is something measurable today that gives you a statistical edge in predicting what an asset's price will do tomorrow, next week, or next quarter. Not a guarantee. Not a prophecy. An edge. A slight tilt of the odds in your favor that, repeated across many bets, compounds into real returns.

Here is a concrete example. The yield curve (the difference between 10-year and 2-year Treasury yields) has inverted before every US recession in the last 60 years. When it goes negative, recession tends to follow within 6-18 months. The yield curve is a signal. It doesn't predict with certainty. It has a 6-18 month error bar on timing. But it shifts the probability distribution of "will there be a recession?" in a direction you can act on.

A critical thing to internalize early: almost all useful signals are weak. A signal that correctly predicts market direction 55% of the time is considered strong. Most academically validated signals operate in the 51-53% range. This sounds useless until you realize that casinos are built on edges smaller than that, and they do fine. The math works because you apply the edge across many bets, over a long period, with proper position sizing. One bet with a 52% edge is a coin flip. A thousand bets with a 52% edge is a business.

02The Taxonomy of Signals

Not all signals work the same way. Understanding the type of signal you're looking at changes how you build it, how you test it, and how you use it. Here are the major categories.

↗

Momentum / Trend Signals

Assets that have been going up tend to continue going up, and vice versa. This is the most robust anomaly in financial markets, documented across asset classes, geographies, and centuries. You measure it by looking at past returns over various windows (1 month, 3 months, 12 months minus the most recent month). The academic literature is deep here: Jegadeesh and Titman (1993) is the foundational paper, and the effect has survived decades of out-of-sample testing.

The mechanism: information diffuses slowly through markets. Not everyone reacts to new information simultaneously. Early movers create a trend; late movers sustain it.

↩

Mean Reversion Signals

Assets that have moved too far from their historical norm tend to snap back. This is the opposite of momentum, and both can be true simultaneously at different time horizons. Short-term (days to weeks), markets tend to overreact and revert. Long-term (months to years), they tend to trend. A z-score (how many standard deviations is the current value from its rolling mean) is the simplest mean reversion signal. When a stock is trading 3 standard deviations below its 200-day average, history says it's more likely to go up than down over the next quarter.

The tension between momentum and mean reversion is one of the central puzzles in quantitative finance. The answer is: they operate on different time scales and in different regimes.

⚡

Event-Driven Signals

Something happens in the world, and you trade the consequence. An earnings surprise, a CEO resignation, a war, a regulatory change, a natural disaster. Event-driven signals are not statistical patterns mined from historical data. They are causal chains: event X causes effect Y, which impacts asset Z. The signal is in recognizing the chain before the market fully prices it.

Event-driven signals are inherently harder to backtest because each event is unique. You can backtest the category of event (oil supply shocks, for example) but not the specific instance (the 2026 Hormuz closure). This is where qualitative judgment meets quantitative rigor.

⚖

Positioning Signals

How other market participants are positioned tells you something about what happens next. When every hedge fund is short crude oil, there's a mechanical asymmetry: if the price goes up even a little, the shorts face losses and may be forced to buy to cover, which pushes the price up further, which forces more covering. This is the short squeeze dynamic, and the positioning data (from CFTC Commitments of Traders reports, or FINRA short interest data) is the signal.

Positioning signals are contrarian at extremes. When everyone agrees on a trade, the trade is crowded, and the next move tends to be in the opposite direction. This has strong academic support: Rapach, Ringgenberg, and Zhou (2016) showed that aggregate short interest predicts equity returns with an out-of-sample R² of 13%.

🏗

Fundamental Signals

Financial statement data — revenue growth, earnings quality, balance sheet health, capital expenditure trajectories — forms the basis for fundamental signals. The idea is that the intrinsic value of a company can be estimated from its financials, and when the market price deviates significantly from that intrinsic value, an opportunity exists.

The challenge with fundamental signals is timing. A company can be fundamentally cheap and stay cheap for years. The signal tells you what to buy but not when. Combining fundamental signals with momentum or event-driven triggers addresses this: buy cheap companies, but only when the price starts moving in your favor.

🌡

Regime / Macro Signals

Instead of predicting the direction of a specific asset, regime signals try to identify the overall market environment: are we in a risk-on or risk-off regime? Is the economy expanding or contracting? Is inflation accelerating or decelerating? The regime determines which other signals work and which don't.

Momentum strategies perform well in trending regimes and poorly during regime transitions. Mean reversion works during stable regimes and fails during structural breaks. A regime signal doesn't tell you what to buy. It tells you which of your other signals to trust right now, and how much exposure to take overall.

↔

Divergence Signals

Two things that normally move together stop moving together. Oil goes up but energy stocks don't follow. The VIX spikes but credit spreads stay calm. Gold crashes during a war. These divergences are information. They tell you that the market is processing conflicting narratives, and one of them is wrong.

Divergence signals require you to have a model (even an informal one) of what "normal" looks like. You need to know that oil and energy stocks are supposed to be correlated before you can recognize when the correlation breaks. This is why breadth of data matters. The more relationships you track, the more divergences you can detect.

🔗

Supply Chain / Second-Order Signals

First-order effects are obvious: Hormuz closes, oil goes up. Second-order effects require reasoning: Hormuz closes, fertilizer exports are disrupted, planting costs increase, farmers reduce acreage, crop yields fall, food prices rise six months later. Second-order signals are where the alpha lives because the market prices first-order effects quickly but takes weeks or months to fully price the downstream consequences.

Building these signals requires domain knowledge and causal reasoning, not just statistical pattern matching. You need to understand how supply chains work, how costs transmit through the economy, and what the time lags are at each step.

03Where Signals Come From

Signals don't fall from the sky. They originate from a specific process that, when done honestly, looks like this:

Observation

You notice something in the data. Crude oil inventories are building while the price is rising. Credit card delinquency was improving for four quarters and just got hit by a gas price shock. Hedge funds are aggressively short a commodity at its 52-week high. The observation is not yet a signal. It's a curiosity that deserves investigation.
Hypothesis

You form a testable claim about cause and effect. "When crude inventories diverge from price, the inventory trend eventually wins." Or: "extreme positioning in CFTC data leads to mean reversion in the following 4-8 weeks." The hypothesis must be specific enough to be wrong. If it can't be falsified, it's not a hypothesis; it's a vibe.
Mechanism

You articulate why the relationship should exist. This is the step most people skip, and it's the most important. A statistical correlation without a causal mechanism is fragile. It might be data mining. It might be coincidence. It might be real but only in a specific regime that has since ended. The mechanism gives you a reason to believe the relationship will persist out of sample.

Example mechanism: "CFTC positioning extremes lead to mean reversion because, at extreme positioning, the incremental buyer/seller is being forced to trade (margin calls, risk limits, fund redemptions), not choosing to trade based on information. Forced trading is uninformed trading, and uninformed flows revert."
Data

Can you actually measure the thing? How far back does the data go? How frequently is it updated? Is it free or expensive? Is it clean or messy? Is there a survivorship bias or a look-ahead bias? The best hypothesis in the world is worthless if you can't observe the inputs reliably.
Feature Engineering

Raw data is almost never useful directly. Oil inventory levels are meaningless. The rate of change of oil inventories relative to the seasonal norm is a feature. The z-score of credit card delinquency relative to its 5-year rolling distribution is a feature. Feature engineering transforms raw observations into comparable, normalized inputs that a signal can use.
Backtesting

You test the hypothesis against historical data. If your signal says "when X crosses above Y, buy SPY," you simulate that strategy over all available history and measure the results. This is where most quantitative work happens, and it's also where most mistakes are made. We'll cover the pitfalls in section 5.
Validation

Statistical significance, out-of-sample testing, robustness checks. Is the result real or random? Does it work across different time periods? Different assets? Different parameter choices? A signal that only works with a lookback of exactly 47 days on US large cap stocks during the third quarter is not a signal. It's an artifact.
Deployment and Monitoring

You put the signal into production. It runs automatically, generates scores, and feeds into a portfolio construction process. And then you watch it. Signals decay. The edge that worked for a decade can stop working when enough people discover it, or when the market structure changes. Monitoring is not optional; it's the difference between a systematic fund and a backtesting exercise.

04How Signals Are Measured

The language of signal quality revolves around a few key metrics. Understanding them is essential to evaluating whether a signal is worth deploying.

Information Coefficient (IC)

The IC is the correlation between your signal's prediction and the actual future outcome. If your signal gives each stock a score, and you rank those scores against the actual subsequent returns, the Spearman rank correlation between the two is your IC.

IC Range	Interpretation	In Practice
0.00 - 0.02	Noise	Can't distinguish from random
0.02 - 0.05	Weak but potentially useful	Needs high breadth to matter
0.05 - 0.10	Solid	Most published academic factors
0.10 - 0.20	Strong	Top-tier institutional signals
> 0.20	Exceptional	Rare. Verify you're not overfitting.

An IC of 0.05 sounds pathetically small. It means your signal explains about 0.25% of the variance in future returns. But the Fundamental Law of Active Management tells you why it's enough.

The Fundamental Law of Active Management

IR = IC × √BR

Information Ratio equals Information Coefficient times the square root of Breadth.

IR is the risk-adjusted return of your strategy. IC is the quality of each prediction. BR is the number of independent predictions you make per year. The square root means that doubling your breadth (number of bets) only increases your IR by about 41%, so you need a lot of bets to make a weak signal work. But this is exactly why quantitative approaches dominate: a computer can make thousands of independent predictions per year across hundreds of assets, and a human can't.

Example: Why breadth matters

Signal A has IC = 0.10 and makes 4 bets per year (quarterly rebalancing on a single asset class).

IR = 0.10 × √4 = 0.10 × 2 = 0.20

Signal B has IC = 0.03 and makes 500 bets per year (daily predictions on 2 assets).

IR = 0.03 × √500 = 0.03 × 22.4 = 0.67

Signal B has a far weaker IC but produces a much higher Information Ratio because of breadth. This is the mathematical foundation for why many weak signals combined correctly outperform fewer strong signals.

Directional Accuracy

The percentage of predictions where your signal correctly identified the direction of the move (up or down). A directional accuracy of 55% over a large sample is very good. An accuracy of 72% at a quarterly horizon, as some regime models achieve, is exceptional. Note that directional accuracy alone is not sufficient. A signal can be right 70% of the time but still lose money if the 30% of wrong predictions produce much larger losses than the 70% of correct predictions produce gains. You need to evaluate both the hit rate and the payoff distribution.

Statistical Significance

A p-value answers the question: "if this signal had no predictive power whatsoever, what's the probability I'd see results this good by chance?" A p-value of 0.05 means there's a 5% chance the result is random. A p-value of 0.001 means 0.1%. In quantitative finance, where you're testing many hypotheses, you need to be stricter than the standard academic threshold. If you test 100 signals, 5 will appear significant at p < 0.05 purely by chance. This is the multiple comparisons problem, and it destroys more backtests than any other methodological issue.

05How Signals Go Wrong

There are more ways to build a bad signal than a good one. The following are not theoretical risks. They are the primary causes of strategy failure in live trading.

Look-Ahead Bias. Using data that wasn't available at the time of the prediction. Q4 2025 GDP data wasn't published until February 2026. If your backtest uses the Q4 GDP number to make a "prediction" in January 2026, you're cheating. The signal looks great in the backtest and fails in production because the real-time data isn't available yet. Every data point has two timestamps: when it occurred and when it was published. The backtest must use the publication date, not the occurrence date.

Overfitting. Tuning your signal to match historical data so precisely that it captures noise rather than signal. If you test 200 parameter combinations and pick the best one, you're not finding a good signal. You're finding the parameter that most closely matches the random fluctuations in your specific dataset. The antidote is out-of-sample testing: hold out a portion of your data that the optimization never sees, and test the signal against it. If performance degrades significantly out of sample, you overfit.

Survivorship Bias. Testing only on assets that exist today. If you backtest a stock-picking strategy using the current S&P 500 constituents, you're excluding every company that went bankrupt, was acquired, or was delisted over the test period. Those are exactly the cases where signals would have been most useful (and possibly most wrong). The data must include dead companies.

Crowding and Alpha Decay. A signal that worked for decades can stop working when enough capital chases it. Once "buy stocks with low P/E ratios" became common knowledge, the returns to value investing compressed. The signal was real; the edge decayed. This is not a backtesting error. It's a structural feature of competitive markets. Any edge that is widely known is no longer an edge.

Regime Dependence. A signal that works beautifully during economic expansions and fails catastrophically during recessions. Or one that worked in a zero-interest-rate environment and breaks when rates rise. If your backtest period doesn't include the regime you're about to enter, your results are misleading. The most dangerous sentence in quantitative finance is "it's never done that before."

06Combining Signals

Single signals are fragile. Combined signals are robust. This is not a theoretical claim. It is a mathematical consequence of the Fundamental Law.

The simplest combination is an equal-weighted average: compute each signal's output, normalize them to the same scale (usually -1 to +1, where -1 is maximally bearish and +1 is maximally bullish), and average them. When five independent signals each say "slightly bearish" and one says "very bullish," the average is "slightly bearish." The dissenting signal might be right and the five might be wrong, but the probabilistic bet is to go with the consensus.

More sophisticated approaches weight signals differently based on their historical reliability. If Signal A has an IC of 0.10 and Signal B has an IC of 0.03, you might give A three times the weight of B. Inverse-variance weighting, decay-weighted combinations, and machine-learning-based meta-models all improve on equal weighting in theory, though the improvement is often modest in practice. The biggest gains come from adding a new independent signal, not from better weighting of existing ones.

The word independent is doing heavy lifting in that sentence. Two signals built from the same data (say, 1-month and 3-month momentum) are correlated. Adding the second one doesn't truly increase your breadth. True breadth comes from combining signals that measure genuinely different phenomena: momentum plus value plus positioning plus macro regime plus sentiment. Each draws on different data, captures different market dynamics, and fails in different environments. The diversification across signal types is analogous to diversification across assets, and it's just as powerful.

07Signals Don't All Look the Same

This is the part that trips up most people building their first systematic strategy. Not every signal is a time series you can backtest across 20 years of clean daily data. Signals come in wildly different shapes:

Some signals are continuous, some are discrete. A z-score is continuous: it can be -2.3 or +0.7 or anything in between. An event signal is discrete: the war either started or it didn't. Both can be useful, but they combine differently. Continuous signals can be averaged. Discrete signals are better modeled as regime switches or conditional filters.

Some signals update daily, some quarterly. VIX updates continuously. GDP updates quarterly with a 30-day lag. If your signal engine requires all inputs to be at the same frequency, you need to decide what to do with the stale inputs. The simplest approach: carry forward the most recent value until a new one arrives. A more nuanced approach: decay the confidence of a signal as its underlying data ages. A 3-week-old CPI reading deserves less weight than a 1-hour-old VIX reading.

Some signals have deep history, some don't. FRED data goes back to the 1940s. CFTC positioning data starts in 2006. Certain alternative data sources only exist from 2020. You cannot build a regime detection model that requires all inputs to have the same history. Instead, the model adapts to what's available at each point in time. In 2005, the model uses macro and price data. In 2007, CFTC positioning becomes available and gets added. In 2020, a new data source appears and the model absorbs it. The model's dimensionality grows over time, and that's fine.

Some signals are stored, some are computed on demand. A z-score of credit card delinquency can be precomputed and stored in a database. But the answer to "what was happening in the world during the closest historical analog to this week?" requires a live query against a massive event database, followed by language model processing and embedding comparison. That's not a stored feature. It's a just-in-time computation. Both are signals; they just have different architectures.

The mistake is to think that because the regime fingerprint approach (put everything in a vector, find historical analogs) works well for macro signals with deep history, it must be the right approach for everything. It isn't. An event-driven signal triggered by a CISA cybersecurity vulnerability disclosure doesn't fit into a regime fingerprint. A short-squeeze signal triggered by FINRA data showing 8+ days to cover doesn't need 20 years of history to be useful. The signal architecture should match the signal type, not the other way around.

08The Practitioner's View

If you're actually building a signal-based system (not just reading about one), here are the things that matter most and are hardest to learn from textbooks:

Start with the data, not the model. Most failed quant projects begin with someone saying "let's build a machine learning model" and then going to look for data to feed it. This is backwards. Start by understanding what data exists, what it measures, how reliable it is, and what relationships you can see in it with your own eyes. The best signals are often simple transformations of well-understood data, not complex models applied to poorly understood data.

The mechanism matters more than the backtest. A backtest tells you what happened. A mechanism tells you why it happened and whether it will keep happening. If you can't explain in plain language why your signal should work, it probably doesn't. The explanation doesn't need to be novel. "Crowded shorts produce mean reversion because forced covering creates uninformed buying pressure" is a perfectly adequate mechanism. It's been known for decades and it still works.

Complexity is not alpha. A 3-feature linear model that you understand completely will outperform a 200-feature neural network that you don't understand, because when the model breaks (and it will), you'll know why and how to fix it. Complexity is a cost, not a benefit. Every additional parameter is an opportunity to overfit. The right question is not "what's the most sophisticated model I can build?" but "what's the simplest model that captures the effect?"

Your signal is only as good as your worst data source. Garbage in, garbage out. A beautifully designed signal that ingests stale data, duplicate records, or data with look-ahead bias will produce confident-looking but worthless outputs. Ninety percent of the work in quantitative finance is data engineering. The modeling is the easy part.

The hardest part is doing nothing. When your signal is neutral, the correct action is to do nothing. When the signal is low-confidence, the correct action is to reduce position size. When you don't understand what the market is doing, the correct action is to go to cash. The temptation to override the signal with intuition is overwhelming, especially during volatile markets. Resisting that temptation is the discipline that separates systematic investors from people who built a system and then ignore it.

Analysis powered by Hypercube Capital's proprietary data aggregation platform and signal engine.

Disclaimer: This material is published by Hypercube Capital for educational purposes only. It does not constitute investment advice or a recommendation to buy or sell any security. All investing involves risk. The concepts described here are simplified for clarity and may not capture the full complexity of real-world implementation.