Detection Is Not Prediction: An Important Distinction in Alternative Data

Executive Summary

Detection and prediction answer different questions.

Detection asks whether something has changed or is changing in the observable world. Prediction asks what will happen next. Alternative data can be useful for both kinds of analysis, but confusing the two creates one of the most common conceptual errors in data-driven investment research.

A satellite observation may detect construction activity. Vessel data may detect changing shipping patterns. Transaction data may reveal a change in spending behavior. None of these observations automatically predicts a future financial outcome.

Prediction requires an additional layer of evidence: a historically supported relationship between the observed condition and the future outcome being forecast.

The distinction matters because a dataset can be highly valuable for situational awareness even when it has little or no demonstrated predictive power.

Key Takeaways

• Detection identifies observable change; prediction estimates a future outcome.

• A signal can be accurate as a detector without being useful as a predictor.

• Prediction requires historical validation beyond the quality of the underlying observation.

• The economic mechanism connecting an observation to an outcome should be explicit.

• Alternative data can provide substantial value even when its primary role is observation rather than forecasting.

What Is Detection?

Detection is the identification of a condition, event or change from available evidence.

Examples include detecting new construction at an industrial site, lower vessel activity around a port, increased congestion along a transport corridor or unusual changes around a physical facility.

The observation may be important.

But detection makes a relatively narrow claim:

Something observable appears to have changed.

It does not necessarily explain why the change occurred. It does not establish its economic significance. And it does not say what will happen next.

That narrower claim is not a weakness. It is a clearer description of what the evidence actually supports.

What Is Prediction?

Prediction attempts to estimate an outcome that has not yet occurred.

Depending on the analytical problem, the target might be future production, future trade volumes, company earnings, commodity balances or financial-market outcomes.

Prediction therefore requires more than observing the current state.

It requires a reason to believe that information available today contains information about a later outcome.

That relationship may be supported by an economic mechanism, historical evidence or both.

Why the Difference Matters

Suppose satellite imagery reveals substantial construction activity at a semiconductor site.

The detection question is:

Has physical construction activity increased?

The prediction question might be:

Will semiconductor production increase next year?

These are not equivalent.

Between construction and future production sit several intermediate steps:

Construction → Equipment installation → Qualification → Production ramp → Utilization → Output

A reliable construction detector could correctly identify expansion while still performing poorly as a predictor of the timing or magnitude of future semiconductor production.

That does not make the detection useless.

It means the claim should match the evidence.

A Detector Can Be Valuable Without Predicting Markets

Financial research often evaluates new data by immediately asking whether they predict returns.

That is understandable, but it is not the only useful standard.

Detection can support monitoring, risk identification, due diligence, event confirmation, operational awareness, supply-chain analysis and company exposure assessment.

Consider a private equity investor monitoring construction at a portfolio company.

The objective may not be to predict a stock price.

The relevant question may simply be:

Is the physical expansion progressing in a way that is consistent with the operating plan?

That is an observational problem.

Accurate detection can be valuable by itself.

Detection Quality and Predictive Quality Are Separate

A useful analytical framework separates two questions.

Is the observation accurate?

This concerns measurement.

Did the system correctly detect the physical or behavioral change?

Does the observation predict the desired outcome?

This concerns historical relationship and economic transmission.

Did similar observations historically contain information about what happened later?

A strong answer to the first question does not imply a strong answer to the second.

For example, a system might detect port congestion reliably. But congestion may have very different downstream consequences depending on duration, cargo type, inventories, alternative ports, shipping schedules and existing market expectations.

Prediction requires understanding those additional conditions.

Observation, Translation and Outcome

Alternative data can be thought of in three distinct stages.

Observation

What physically or behaviorally changed?

Translation

What might that change mean economically?

Outcome

What happened later?

The first stage is primarily a measurement problem.

The second is an interpretation problem.

The third creates an opportunity for historical validation.

Collapsing all three stages into a single concept called a “signal” can obscure important uncertainty.

What Can Be Observed?

Alternative datasets can directly reveal conditions represented by their measurements.

Satellite data may reveal physical surface changes.

AIS may reveal vessel movements.

Mobility data may reveal observed movement patterns.

Transaction datasets may reveal spending captured within their coverage.

These are observations.

Their strength should first be evaluated on whether they accurately represent what they claim to measure.

What Can Reasonably Be Inferred?

An observation can support an interpretation when there is a credible mechanism and sufficient context.

For example:

Sustained lower vessel departures from a known export terminal may be consistent with reduced outbound shipping activity.

That inference remains relatively close to the observation.

A stronger claim such as:

The country's exports will decline next month

requires considerably more evidence.

The distance between observation and conclusion matters.

What Cannot Be Concluded From Detection Alone?

Detection alone generally cannot establish:

• future earnings,

• future asset prices,

• causal economic effects,

• exact production levels,

• financial materiality,

• market surprise.

A physical change can be real without being economically important.

An economically important change can be real without being unexpected.

And even unexpected information does not determine how an asset will trade.

Why Historical Validation Matters

Prediction should be tested against outcomes that occur after the signal.

A disciplined historical evaluation asks questions such as:

• Was the observation available at the time?

• Was the hypothesis defined independently of the result?

• What happened afterward?

• How often did the proposed relationship hold?

• Did it persist across different periods?

• Did performance depend on a particular threshold or horizon?

• Was the effect economically meaningful?

Backtesting can help approximate how a strategy or analytical rule would have behaved historically.

But historical tests can also be distorted by look-ahead bias, survivorship bias, overfitting and structural change.

A prediction should therefore be treated as an empirical claim, not as a semantic upgrade from detection.

A Hypothetical Example

Consider an industrial facility monitored using Earth observation.

During several weeks, observable activity around the site falls sharply.

Detection

The activity measure declines.

Interpretation

The change may be consistent with lower facility activity.

Company Exposure

A company operates the facility and therefore has direct operational exposure.

Prediction Hypothesis

The decline may precede weaker reported production or financial results.

Validation

Historical observations must then be compared with later outcomes.

Only this final step can establish whether the proposed predictive relationship has empirical support.

Without validation, the prediction remains a hypothesis.

Why Prediction Is Harder Than Detection

Prediction introduces uncertainty at multiple layers.

First, the observation itself can be noisy.

Second, the economic interpretation may be wrong.

Third, the transmission mechanism may change.

Fourth, the target outcome may depend on many unrelated variables.

Finally, financial markets incorporate expectations from many sources.

Every additional layer introduces another way for the relationship to fail.

This is why a system that reliably describes the present does not automatically predict the future.

The Value of Knowing the Current State

There is sometimes an unnecessary hierarchy in investment research where prediction is viewed as sophisticated and observation as merely descriptive.

In practice, accurately understanding the current state of a physical or economic system can be extremely difficult.

Official statistics are often delayed.

Company reports summarize periods that have already ended.

Physical-world data can provide additional evidence about activity occurring between those reporting events.

A better view is therefore:

Observation and prediction solve different problems.

One does not have to become the other to be useful.

Space Sat Lab Perspective

Space Sat Lab separates physical observation from subsequent interpretation and historical evaluation.

Within Planetary Economic Observability, the first objective is to establish what has actually changed in the observable world.

Economic transmission can then be examined to understand potential relevance. Historical outcomes can finally be used to test whether particular observations have contained repeatable information about later events.

That ordering helps prevent a useful observation from being promoted into a predictive claim before the evidence supports it.

Reality first.

Prediction, if justified, comes later.

Frequently Asked Questions

Is alternative data predictive?

Some alternative datasets can contain predictive information, but predictive value must be demonstrated empirically. Capturing activity earlier or differently from traditional sources does not automatically make a dataset predictive.

What is the difference between a signal and a prediction?

A signal is evidence or information that may be analytically relevant. A prediction is a claim about a future outcome. A signal can exist without having demonstrated predictive power.

Can satellite data predict company earnings?

Satellite observations may contain information related to company operations, but a relationship with future earnings requires separate historical validation and a credible economic transmission mechanism.

Does accurate detection make a dataset useful?

It can. Detection can support monitoring, risk analysis, due diligence and situational awareness even when prediction is not the objective.

Why is backtesting important?

Backtesting allows researchers to examine how a proposed rule or relationship would have behaved historically. Results still require careful interpretation because historical tests can be affected by bias, overfitting and changes in underlying relationships.

Sources and Further Reading

CFA Institute, Backtesting & Simulation.

CFA Institute, research and educational material on look-ahead bias and investment model validation.