The alternative data industry grew from a handful of specialist providers to a market worth several billion dollars in under a decade. For systematic investors, the key question is not whether alternative data exists but which of it actually generates durable alpha — and why most of it doesn’t.
The Signal-to-Noise Problem
Alternative data ranges across categories: consumer transaction data, satellite imagery, shipping and logistics signals, web scraping, employment data, and natural language processing of corporate communications. Each category was, at some point, genuinely differentiated information.
The market dynamics of alpha decay are relentless. When a signal works — when it demonstrably predicts price movement — capital follows. As capital follows, the market incorporates the signal faster and the edge compresses. What was a ten-day return window becomes a two-day window becomes intraday and then disappears.
The implication for systematic investors is that the evaluation question is not “does this data source work?” in isolation, but “when does it stop working, for whom, and at what scale?”
Where the Real Signal Lives
The alternative data that has shown the most durable alpha characteristics is not the most exotic. It is often the most public — corporate language in regulated communications that everyone can access, but very few can process at the depth required to extract genuine signal.
Earnings calls, annual filings, and regulatory disclosures contain structured information that companies produce on predictable schedules. Unlike transaction data or satellite imagery — which can be acquired by anyone with sufficient budget — the signal here is not in the raw data. It is in the calibration.
Raw sentiment scores derived from corporate language are commoditised. Anyone can buy that. The signal that generates returns is in the longitudinal interpretation: not what a company is saying today in isolation, but how that compares to what it has said historically — and what a meaningful change in that pattern signals about the underlying business.
That calibration takes years of data and significant engineering to build. Which is precisely why it remains available.
The Point-in-Time Discipline
The most common failure mode in alternative data backtests is survivorship and look-ahead contamination.
Data vendors, when they onboard a new client, typically provide a historical dataset covering several years. That dataset was constructed with the benefit of knowing which companies survived, which data was ultimately collected, and which gaps were filled in retrospectively. A backtest on that data overstates performance because it uses information that was not available at the time the hypothetical trade would have been made.
Rigorous systematic investing requires point-in-time correct data: the exact information set that was available as of each historical date, including gaps, delays, and restatements. Building and maintaining a point-in-time dataset is operationally expensive. Most managers cut this corner. It is why many backtests look significantly better than live performance.
What Survives Validation
Most alternative data signals fail at least one of the following tests:
Out-of-sample decay. The signal worked on the training dataset but shows rapid decay on the validation period — a clear sign of in-sample overfitting.
Correlation with existing factors. The signal appears to predict returns, but when factor-adjusted, the apparent alpha disappears. It was a correlated proxy for a known factor rather than an independent source of information.
Cost-adjusted returns. Transaction costs, data acquisition costs, and market impact consume the gross alpha. This is common in high-frequency signals applied to mid-cap universes.
Regime instability. The signal worked in one market environment but underperformed or reversed in others. Without understanding why the signal works mechanically, it is impossible to know whether historical outperformance reflects a genuine relationship or a coincidence of conditions.
The managers who maintain this discipline consistently — and reject the majority of what they evaluate — are the ones most likely to hold signals that survive real markets.
The Compounding Advantage
The durable edge in systematic investing is not in having access to better data. It is in having a more disciplined and comprehensive process for evaluating what that data means.
Any sophisticated manager can buy the same datasets. The differentiation is in the validation framework, the point-in-time discipline, and the accumulated calibration that transforms commoditised data feeds into proprietary signal. That calibration cannot be purchased. It has to be built.
AvB Capital is an AI-native systematic manager based in London. Track record and strategy documentation are available to qualified investors on request.
Frequently Asked Questions
What counts as alternative data in investing?
Alternative data is any non-traditional information source used to inform investment decisions — satellite imagery of retail car parks, credit card transaction data, shipping container movements, job posting volumes, social media sentiment, and increasingly, large-scale processing of corporate language in filings, earnings calls, and press releases. The defining characteristic is that it sits outside the standard financial data universe that all market participants access.
Does alternative data actually generate alpha?
Some does. The majority does not — or does not at the cost of acquiring it. The critical factor is whether the data captures something the market has not yet priced. Most heavily marketed alternative data sources are already embedded in price by the time smaller managers access them. The signal, when it exists, is typically in the second-order interpretation: not the data itself, but the change in pattern relative to an established baseline.
How does natural language processing extract signals from corporate filings?
LLMs process the text of earnings calls, annual filings, and regulatory disclosures to extract structured signals: tone, specificity of forward guidance, mention frequency of risk factors, comparison to prior periods. The raw sentiment score is typically commoditised. The valuable signal comes from comparing the current reading to that company's own historical baseline — detecting when something has shifted, not just whether it is positive or negative.
What is the survivorship bias problem in alternative data backtests?
Most alternative data providers backfill historical data at the time of onboarding, which means the dataset looks clean and complete in hindsight. In reality, data coverage was irregular, delayed, or missing for many companies during the historical period. A backtest on backfilled data overstates expected returns because it includes the benefit of information that was not actually available at the time of the hypothetical trade.
How do systematic managers evaluate whether alternative data is worth using?
The evaluation requires out-of-sample testing with point-in-time correct data, decay curve analysis, correlation analysis with existing signals, and cost-adjusted return calculation. The majority of evaluated datasets fail one or more of these tests.