Originally published July 2022. Revised and updated August 2026.
There is no shortage of writing about what alternative data is. There is much less about what it produces when someone actually builds a strategy on it and tests the result.
Below are three quantitative strategies built on three unrelated alternative datasets — search behaviour, media sentiment, and government contract receivables. Two are published by third parties. One is ours.
They have almost nothing in common at the data level. But they share a construction principle, and it is the single most useful thing to understand before building a signal on any alternative dataset.
The common thread: levels are not signals
None of the three strategies below trades on the raw value of its dataset. Each one transforms the data into a measure of deviation: how far the current reading sits from what would have been expected for that particular security.
There is a straightforward reason this keeps recurring. Raw alternative data levels are dominated by size. Large companies attract more search traffic, more news coverage, more shipments and more government contracts than small ones, more or less mechanically. A signal built on unnormalised levels will sort stocks largely by market capitalisation, backtest respectably for reasons that have nothing to do with the dataset, and fail the moment it is controlled for size.
The work in building an alternative data signal is rarely in acquiring the data. It is in deciding what “normal” looks like for each security, and measuring the departure from it.
Strategy 1: Search attention
Quantpedia, which maintains a database of quantitative strategies drawn from academic research, catalogues a strategy built on Google search volume.
The construction is deliberately contrarian: buy stocks attracting low search attention, and sell as search traffic picks up. The underlying premise is that low-attention stocks are systematically underpriced and carry lower risk, and that rising retail attention marks the point at which that discount closes.
Note what the signal is not. It is not “companies people search for often” — that list would be Apple, Tesla and Amazon in perpetuity. It is the change in attention relative to a stock’s own norm. The strategy is built on the first derivative, not the level.
Strategy 2: Media sentiment
Refinitiv (now part of LSEG) published a white paper testing media sentiment as a standalone investment signal against both a traditional multi-factor strategy and the S&P 500.
The results were closer than one might expect. Used as a single-factor strategy, media sentiment produced returns, volatility and Sharpe ratios broadly comparable to a multi-factor approach. For long-only portfolios, both the multi-factor and sentiment-only strategies delivered between 2% and 2.5% excess return over the benchmark.
The more interesting result appeared in long-short portfolios, where the sentiment-only strategy delivered 3% to 5.5% excess return and began to outperform the multi-factor model outright. The effect strengthened as the short allocation increased — portfolios with 50% or more in short positions saw the sentiment signal beat both the benchmark and the multi-factor strategy.
That asymmetry is worth pausing on. It suggests sentiment carries more information about deterioration than about improvement, which is consistent with how negative news propagates relative to positive news.
Sentiment scoring conventionally measures tone relative to a security’s own coverage history rather than on an absolute scale, though the white paper does not detail the normalisation. Of the three strategies here, this is the one where the common thread is inferred rather than documented.
Strategy 3: Unexpected government receivables
Government procurement data is public, but it is fragmented across agencies and jurisdictions, published in inconsistent formats, and expensive to process. That combination is what makes it interesting: the information is available, and largely unpriced.
We built a signal on it. Using forward-looking receivables from US federal contracting, the unexpected government receivables (UGR) measure is constructed in two steps: scale each stock’s monthly government receivables by its market capitalisation, then normalise that figure against the firm’s own trailing twelve-month mean and standard deviation.
The second step is the one that matters, and it is the clearest illustration of the principle above. A defence prime winning a large federal contract is not news — it is what a defence prime does. The signal is designed to fire only when a company’s contracted future receivables depart from its own established pattern.
Backtested across US-listed equities, strategies built on UGR produced:
- Annualised Sharpe ratios between 0.77 and 1.27, against a historical market portfolio Sharpe of 0.35–0.40
- Alphas between 3.4% and 7.1% per year, significant across CAPM and multiple Fama-French specifications
- Alpha persisting up to seven months before losing statistical significance
A subsequent case study applied the same data to two concentrated portfolios — ten aerospace and defence suppliers benchmarked against the SPDR S&P Aerospace & Defense ETF, and thirteen large contractors across IT, industrials, construction, healthcare, insurance and services benchmarked against the S&P 500. Both produced positive average alpha per trade.
The usual caveats apply and should not be skipped: these are historical backtests on a US-listed universe over defined sample periods, and they have not been re-tested through the fiscal disruptions of the period since.
Full methodology, factor models and results →
What this means if you are building your own
Four practical consequences follow from the common thread.
Decide what “expected” means before you look at returns. A trailing mean over the security’s own history is the simplest baseline and often adequate. Industry-relative and size-relative baselines are alternatives. What matters is choosing the baseline on reasoning rather than on which one backtests best — the latter is how overfitting starts.
Scale before you normalise. UGR scales by market capitalisation first, then normalises. Order matters: normalising an unscaled series leaves the size effect intact inside the deviation.
Check whether your signal survives a size control. If a factor model that includes size absorbs most of your alpha, the dataset was not telling you anything. This is the fastest diagnostic available and it should be run early, not last.
Beware baselines that need long history. A twelve-month trailing baseline consumes a year of data before generating its first observation. On a dataset with three years of history, that leaves two — not enough for a credible test. History requirements compound in signal construction in a way people routinely underestimate.
How much is being spent on this?
A brief note, because the figures circulate widely and are less solid than they look.
Estimates of the alternative data market vary by roughly sixfold depending on the source, from around $2.8bn to $18.8bn for comparable periods. The divergence is definitional — broader estimates include corporate buyers, infrastructure and processing platforms, while narrower ones count only investment-manager spend on datasets.
The direction is consistent across every source: spending is rising quickly. The level is not knowable to the precision usually implied. Treat any single headline figure, including ones we might quote, as directional.
Does the advantage erode?
The standard concern is that any dataset generating returns will be competed away as adoption spreads.
The evidence is more nuanced. Neudata’s 2026 market analysis found the average alternative dataset is used by just 20 investment firms, down from 25 the year before — datasets are becoming more exclusive, not less, because the catalogue is growing faster than the buyer base.
Crowding, in other words, is a property of particular datasets rather than of the category. Card transaction panels covering large US retailers are heavily subscribed. Datasets requiring substantial processing before they yield anything usable are frequently not.
Which points at an uncomfortable but useful conclusion: the datasets most likely to retain an edge are the ones that are annoying to work with. The friction that keeps a dataset from being widely adopted is the same friction that keeps its information from being priced.
Two things that matter more than dataset choice
Combining sources. Few strategies rest on a single dataset. The usual approach is to require confirmation across two or more, which reduces the risk that a signal is an artefact of one provider’s collection methodology.
Data quality, specifically point-in-time integrity. A backtest built on a current snapshot assumes the researcher knew, on each historical date, what the database says today. Where the underlying records have been revised since — and in most real-world datasets they have — the test is using information that was not available at the time. Signals built this way look excellent and fail in production.
This is the question to put to any provider before anything else: can you reproduce your data as it stood on an arbitrary past date? If the answer is no, no backtest on that data means very much. Our own field-level coverage is set out in the data documentation, and the broader landscape in our overview of how hedge funds are using alternative data.
Frequently asked questions
What quantitative strategies use alternative data?
Published examples include strategies built on search volume (trading changes in retail attention), media sentiment (particularly effective in long-short portfolios), and government contract receivables (trading deviations in a firm’s contracted future income from public bodies). Others in common use draw on card transactions, satellite imagery, web traffic, app usage and shipping records.
How do you turn alternative data into a trading signal?
Rarely by using raw values. The standard approach is to establish a baseline expectation for each security — usually from its own trailing history — and measure deviation from it. Raw levels tend to be dominated by company size, so an unnormalised signal often functions as a size factor rather than as information.
Why normalise alternative data before backtesting?
Because large companies generate more of almost everything: more search traffic, more news, more transactions, more contracts. Without scaling and normalisation, a signal sorts stocks by size. If a factor model including size absorbs most of the alpha, the dataset contributed nothing.
Do alternative data strategies actually work?
Published backtests report meaningful excess returns across several datasets, including Sharpe ratios of 0.77 to 1.27 and annual alphas of 3.4% to 7.1% for government receivables strategies. These are historical results on defined samples, not guarantees. The more useful question about any dataset is what has actually been demonstrated with it, rather than what it could theoretically show.
Does alternative data lose its edge as more funds adopt it?
Crowding appears to be dataset-specific rather than category-wide. Industry analysis in 2026 found the average dataset used by 20 investment firms, down from 25 the previous year, as the number of available datasets grew faster than the buyer base. Datasets requiring significant processing tend to remain less crowded.
TenderAlpha provides government contracting, trade flows and supply chain datasets to institutional investors, mapped to tradeable securities with point-in-time history from 2010.
If you are testing whether procurement data adds signal to an existing process, tell us the universe you cover and we will show you what it looks like for your names.