This is the 6th chapter of the 11th 'Behind The Cloud' series:
The Data Engine - How AI Funds Sense Markets
Omphalos’ long-term development has reinforced one lesson: in live markets, robustness beats cleverness.
Data is where robustness begins.
This series continues the Behind The Cloud mission: to share research-based insights into what truly drives AI investing, beyond buzzwords, beyond demos, and always grounded in real-world constraints.
Trust good (!) data, not just AI.
Alternative Data - Edge or Trap?
It is also one of the least understood.
Satellite images. Web traffic. App usage. Shipping data. Job postings. Credit card panels. Social media. News flow. Search trends. Supply chain signals. Geolocation data. Scraped text. Specialist vendor feeds.
The promise sounds simple: traditional data is crowded, so look somewhere else.
But in live markets, alternative data is rarely simple. It can be expensive, fragile, biased, inconsistent, difficult to validate, and easy to misunderstand. It can look powerful in a backtest and weak in production. It can describe something real while still failing to become an investable signal. It can add context, but also create false confidence.
The question is therefore not whether alternative data is useful. The question is whether it is usable.
In AI investing, alternative data should not be treated as a magic source of truth. It should be treated as another sensor in the data engine, valuable when it improves understanding, dangerous when it is mistaken for certainty.
The Attraction of Alternative Data
The attraction is obvious.
Markets are competitive. Public prices, reported fundamentals, and standard macro indicators are widely available. If everyone looks at the same information, the search for edge naturally moves toward less conventional sources.
Alternative data appears to offer earlier signals.
A change in web traffic before earnings. A shift in hiring patterns before revenue growth changes. Satellite images that hint at activity levels. Supply chain data that reveals stress before it reaches financial statements. Text signals that capture sentiment before analysts update forecasts.
For AI systems, this is tempting because machine learning can process large, messy, high-dimensional datasets. It can identify weak patterns across many sources. It can combine alternative inputs with traditional market data to build richer context.
But temptation is not evidence.
The more unusual the dataset, the more careful the validation must be. A dataset can be interesting, proprietary, and expensive, while still not producing a robust investment signal.
Novelty alone is not edge.
Why Backtests Often Look Better Than Reality
Alternative data often looks impressive in backtests. There are several reasons for this.
First, the data history is often short. Many alternative datasets did not exist ten or fifteen years ago, or existed only in a different form. A short history makes it difficult to test across regimes. It also increases the risk that a result is specific to one market environment.
Second, coverage changes over time. Vendors expand sources, improve collection, add geographies, change filters, and repair historical gaps. What looks like a continuous dataset may actually be a sequence of changing measurement systems.
Third, alternative data is often cleaned after the fact. Missing values are filled. Errors are corrected. Classifications are improved. Historical mappings are rebuilt. That may make the dataset more useful today, but it can make historical testing unrealistically clean.
Fourth, the signal may be discovered after looking at many possible relationships. If enough datasets, transformations, lags, and universes are tested, something will appear predictive. That does not mean it will survive live trading.
It would be good to also make some regression checks or generalization for the market regime period. If data are totally different in similar market regimes, their value is usually very low.
This is why alternative data must be tested with the same point-in-time discipline as traditional data, and sometimes with even more scepticism.
The dataset may be new, but the old dangers remain: overfitting, look-ahead bias, survivorship bias, and hidden revisions.
Future Leakage, When the Future Enters Through the Side Door
One of the most dangerous risks in alternative data is proxy leakage.
Proxy leakage occurs when a dataset appears to provide an early signal, but actually contains information that would not have been available at the decision time, or is indirectly contaminated by later knowledge.
This can happen in subtle ways.
A vendor may classify companies using today’s business descriptions and apply those classifications historically. A dataset may include entities only after they became visible enough to be collected. Survivorship bias must also be considered, not only in stock universes, but across alternative datasets such as store locations, app activity, web traffic, suppliers, products, or customer panels. If closed stores, failed companies, discontinued apps, inactive suppliers, or lost customers disappear from the history, the dataset becomes biased toward what survived. A web signal may be mapped to companies using a later corporate structure. A supply chain dataset may be cleaned using relationships discovered after the event. A sentiment score may be generated with a model trained on future language patterns.
None of this has to be intentional.
But the effect is the same. The past becomes easier to predict because it has been organized with knowledge from the future.
For AI systems, proxy leakage is especially dangerous because the model does not know why the pattern exists. It only sees that the pattern works. It may treat contaminated structure as genuine information.
This is how a backtest becomes convincing and wrong at the same time.
A robust data engine must therefore ask not only whether an alternative dataset predicts something, but whether it could have predicted it at the time.
Novelty Decays
Even when alternative data works, the edge may not last.
Once a dataset becomes known, sold broadly, or incorporated into common investment workflows, its informational advantage can decay. More investors observe the same signal. More models react to it. Prices adjust faster. The correlation that once looked strong becomes weaker. The timing advantage compresses.
This does not mean the data becomes useless. It means the role of the data changes.
A dataset that once produced alpha may become a context variable. It may still help identify regimes, confirm or challenge other signals, improve risk control, or detect crowding. But it may no longer support a standalone trading signal.
This is a common pattern in systematic investing.
Research discovers a relationship. Capital follows. The relationship weakens. What remains is not always nothing. Sometimes what remains is a better understanding of market structure.
Alternative data should therefore be evaluated dynamically. The question is not only: did it work historically?
The better question is: what role does it still play now?
Correlation Is Not an Investable Signal
A common mistake is to confuse correlation with investability.
A dataset may correlate with revenue growth, inflation, consumer demand, supply chain activity, or asset returns. That is interesting. But it is not enough.
The next question is causality, or at least economic logic. Why does the relationship exist? What is the mechanism? Does the data contain information the market has not yet fully priced, or does it merely move after the market has already moved?
This distinction is critical. If alternative data only follows the stock price, it is usually not useful as an investment signal. It may describe what the market has already understood. What matters is the opposite direction: data that the market may follow, because it captures a real-world change before that change is fully reflected in prices.
To become investable, the signal must satisfy harder conditions.
It must be available on time. It must be stable enough across regimes. It must survive transaction costs. It must be linked to instruments that can be traded efficiently. It must add information beyond existing signals. It must have enough history to evaluate. It must be robust to vendor changes. It must not be too crowded. It must not create compliance or ethical problems.
A beautiful correlation that cannot be traded is not an edge. It is research decoration.
This is why alternative data should be evaluated through an investment lens, not only a statistical lens. The system must understand how the signal would enter the portfolio, what exposure it creates, how it interacts with other signals, and when it should be ignored.
Investability is where many alternative data projects fail.
Not because the data is meaningless, but because the path from data to controlled exposure is weak.
Context May Be More Valuable Than Alpha
The biggest value of alternative data is often not a clean alpha signal.
It is context.
Alternative data can help the system understand whether a market move is supported by real activity or driven by positioning. It can help detect when reported fundamentals are lagging current conditions. It can reveal stress in supply chains, shifts in consumer behavior, changes in liquidity demand, or early signs of regime transition.
In this role, alternative data improves sensing quality.
It may reduce false confidence. It may help determine whether a signal deserves more or less trust. It may improve risk controls by identifying environments where normal relationships are becoming unstable.
This is less glamorous than finding a hidden alpha source. But it may be more durable.
A dataset that helps the system avoid bad exposure can be just as valuable as one that helps identify good exposure. In autonomous investing, improving when not to trade, when to reduce size, or when to require confirmation can be a significant advantage.
Alternative data is therefore often best understood as part of the sensing engine, not as a standalone prediction machine.
Fragility and Cost
Alternative data can be expensive. But cost is not only the subscription price.
There is integration cost. Cleaning cost. Storage cost. Legal review. Vendor due diligence. Monitoring. Governance. Historical validation. Methodology tracking. Model maintenance. Change control. Compliance. Operational dependency.
There is also fragility.
Some alternative datasets depend on collection methods that can change or disappear. Websites block scraping. Platforms change APIs. Privacy rules evolve. Vendors lose access. Panels become less representative. Data providers change their methodology. Coverage improves in ways that break historical comparability. A signal can decay because the world changed, or because the dataset changed.
This makes alternative data a supply chain problem.
A fund must know not only what the data says, but how the data is produced, how stable that production process is, and what happens if it changes. Without that, the dataset becomes a hidden dependency.
In AI investing, hidden dependencies are risk.
Governance and Ethics
Alternative data raises governance questions more sharply than traditional market data.
Where did the data come from? Was it collected lawfully? Is consent clear? Does it contain personal information? Could it create privacy concerns? Is it materially non-public information? Are there jurisdictional restrictions? Does the vendor have the right to sell it? Can the fund explain how it is used? Can it be audited?
These are not legal afterthoughts. They affect whether the data can be used at all.
For institutional AI investing, ethical and governance standards matter because autonomy increases the distance between data source and portfolio action. If a system ingests questionable data and converts it into trades, the issue is no longer theoretical. It has entered the investment process.
A robust data engine therefore needs approval workflows, source documentation, usage restrictions, monitoring, and audit trails. The goal is not only to find information others do not have. The goal is to use information responsibly and defensibly.
In alternative data, governance is part of edge. Without it, the edge may not be usable.
When Alternative Data Becomes a Trap
Alternative data becomes a trap when it is treated as truth.
It becomes a trap when a short history is mistaken for robustness. When cleaned history is mistaken for live reality. When correlation is mistaken for causation. When a proxy is mistaken for direct measurement. When novelty is mistaken for durability. When data availability is mistaken for legal or ethical usability.
It also becomes a trap when it is allowed to overcomplicate the system.
More data can create more features, more models, more explanations, and more opportunities to overfit. A system can become impressive and fragile at the same time. It can become harder to understand which inputs matter, harder to monitor failures, and harder to explain why the portfolio behaves as it does.
This is why alternative data must earn its place in the system.
It should improve the quality of sensing, the robustness of decisions, or the control of risk. If it does not do that, it is complexity.
And complexity has a cost.
Omphalos Perspective
At Omphalos, we do not view alternative data as a shortcut to alpha.
We view it as a potential sensor layer.
That distinction matters. A sensor does not automatically create a trade. It improves the system’s view of reality when it is reliable, timely, validated, and relevant. It can support regime detection, cross-check other inputs, identify stress, and improve confidence estimates. But it can also mislead if it is unstable, contaminated, or poorly governed.
This is why alternative data must pass a higher bar.
We look not only at whether a dataset appears predictive, but whether it is point-in-time, methodologically stable, legally usable, operationally monitorable, and genuinely additive to the existing sensing engine. We also ask whether the signal remains useful once costs, liquidity, crowding, and implementation constraints are considered.
For us, the best alternative data is not the most exotic. It is the data that makes the system more robust.
Key Takeaway
Alternative data is not a shortcut to alpha.
It is a tool that can strengthen a sensing engine, but only if it is validated, governable, and robust. Its value is not measured by how unusual it sounds, or how impressive it looks in a backtest. Its value is measured by whether it improves decision quality in live markets.
In AI investing, more data is not automatically better.
Better sensing is better.
Trust good (!) data, not just AI.
Supporting research & news
Next week we will publish the sixth chapter of this series: "The Cost of Data - When Edge Becomes Fragility'
If you missed our former editions of "Behind The Cloud", please check out our BLOG.
Omphalos Fund won the "Funds Europe Awards 2025" in the category "European Thought Leader of the Year".
Omphalos Fund won the "EuroHedge Awards 2025"
© The Omphalos AI Research Team - September 2026
If you would like to use our content please contact press@omphalosfund.com