Alternative Data for Stock Research

Alternative data for stock research includes information outside the traditional financial statements and market-price series used in fundamental and quantitative analysis. Public social conversation, web activity, app usage, transactions, geospatial observations, and other datasets can improve discovery or measurement. They also introduce serious questions about legality, provenance, bias, stability, and whether the signal adds anything after costs.

What is alternative data?

Traditional equity research commonly uses financial statements, company guidance, market prices, industry data, and macroeconomic releases. Alternative data is a broad label for other information used to estimate business activity, expectations, risk, or investor behaviour. The category describes the source, not the quality.

Examples include aggregated transactions, job postings, web traffic, app rankings, shipping or location observations, patent records, product reviews, public news and social conversation. Some datasets measure business operations; others measure attention or expectations. Treating them as one interchangeable asset class creates bad models and weak diligence.

Where social intelligence fits

Social sentiment, mindshare, narratives, and momentum are forms of public-information market intelligence. They are strongest when the research question concerns expectations, attention, disagreement, event reaction, or narrative diffusion. They are weaker as direct substitutes for revenue, cash flow, market share, or valuation.

Academic research, including Paul Tetlock's work on media content and stock-market activity, has examined relationships between pessimistic language, prices, and trading volume. Such findings do not mean every modern sentiment score predicts returns. Samples, sources, market structure, model design, costs, and publication effects matter. Treat prior research as motivation for careful testing, not a universal trading rule.

Step 1: specify the research job

“Use alternative data” is not a strategy. Define the decision and target variable. Are you trying to discover an earnings catalyst, estimate product demand, monitor reputational risk, measure investor attention, or rank stocks for further research? Each requires different data and evaluation.

Write the mechanism before collecting the dataset. If rising job postings are supposed to anticipate growth, identify which roles, regions, lag, and business segment matter. If stock mindshare is meant to detect narrative diffusion, define the stock universe, sources, and minimum activity. This prevents a convenient metric from becoming a post-hoc explanation.

Step 2: perform data-source diligence

The SEC's examination staff has highlighted policies and diligence around alternative-data providers, including risks involving material nonpublic information. A research organisation should understand how data was obtained, whether collection and use are authorised, which contracts and privacy rules apply, and how the provider monitors changes in provenance.

Ask for the collection method, licensing chain, update frequency, geographic coverage, historical availability, correction process, and known exclusions. Determine whether users can inspect evidence or only receive an opaque score. Review whether the dataset contains personal information or could expose confidential business information. Legal and compliance review is part of the product, not an administrative afterthought.

Step 3: audit coverage and bias

Every source observes a subset of reality. Social platforms overrepresent people who post publicly; app panels may skew toward particular devices or regions; transactions may cover selected merchants; satellite data can be affected by weather and classification error. Coverage can also change when a platform, provider, or collection rule changes.

Measure missingness, duplication, subject ambiguity, language coverage, survivorship, and historical revisions. Check whether the signal is mechanically larger for famous stocks because they receive more data. Use minimum sample thresholds and compare a stock with its own history as well as peers.

Step 4: validate the signal

Divide discovery, model development, and evaluation periods. Avoid building a rule on the same events used to demonstrate it. Include delisted or failed equities where relevant, account for publication timestamps, and remove information that would not have been available at the decision time.

Test incremental value. Does social momentum improve a catalyst screen beyond news volume and price? Does a transaction series add information beyond company guidance and industry releases? Compare the alternative dataset with a simple baseline after realistic latency, coverage, and cost. Statistical significance without an operationally useful effect is not enough.

Step 5: preserve lineage and reproducibility

Store the dataset version, query, subject identity, timestamp, window, transformations, exclusions, and output. If an API value changes after a correction, the research record should reveal which version informed the decision. This is especially important for AI agents, which can blend observations and inferences into fluent but difficult-to-audit text.

A production alert should link to evidence and state uncertainty. A chart should identify units and coverage. A model feature should have an owner and monitoring rule. When source availability, schema, or distribution changes, treat it as a potential break rather than silently extending the historical series.

A practical evidence stack

Example: researching an earnings change

Suppose social momentum around Amazon rises before results and discussion focuses on AWS demand. First verify whether the activity is broad and whether any public catalyst explains it. Record what the posts actually claim. Compare with prior guidance, public industry data, and consensus. After the release, check the claim against reported AWS growth and management commentary.

The alternative signal is useful if it identified a research question earlier or measured a change more clearly. It is not validated merely because Amazon moved. Review false positives and quiet events, not just memorable successes. The earnings sentiment workflow provides a structured event timeline.

Alternative-data checklist

• What exact research variable is the dataset supposed to measure?
• How was it collected, licensed, transformed, and corrected?
• Could it contain confidential or material nonpublic information?
• Which populations, regions, subjects, or periods are missing?
• Can the result be reproduced using information available at the time?
• Does it add value beyond a simple traditional-data baseline?
• How will schema, coverage, and distribution changes be detected?
• Can a reviewer inspect the evidence behind an important conclusion?

Sources and further reading

• SEC Division of Examinations risk alert covering alternative data
• FINRA overview of social-media-influenced investing
• Paul Tetlock: Giving Content to Investor Sentiment

Common questions about alternative data

Is public social data automatically safe to use?

Public visibility does not settle licensing, privacy, platform-terms, intellectual-property, or compliance questions. Collection method, retention, transformation, redistribution, and intended use all matter. Organisations should conduct legal and compliance review appropriate to their jurisdiction and workflow rather than infer permission from the ability to view a page.

What makes an alternative dataset valuable?

Value comes from measuring a relevant variable with enough coverage, timeliness, stability, and incremental information to improve a real decision. A novel dataset that duplicates public filings late, changes methodology without notice, or cannot be mapped to a business mechanism may have little research value despite looking sophisticated.

How should researchers handle a vendor methodology change?

Record the change date, preserve the old definition, test overlap where possible, and avoid presenting the joined series as continuous without qualification. Revalidate models and alerts because a distribution shift can look like a market event. The vendor's correction and version history should be part of the diligence record.

Should alternative data be used alone?

Usually it is strongest as one layer in an evidence stack. Social intelligence can reveal a changing expectation; filings establish reported facts; official datasets provide external context; market data shows price and risk. Independent layers reduce the chance that one biased or broken dataset dictates the conclusion.

How much history is enough?

Enough history must cover the market regimes, seasonality, corporate events, and methodology versions relevant to the intended decision. A short clean series may support operational monitoring but not a claim of durable predictive value. Prefer honest scope limits to backfilling data with rules that could not have been applied contemporaneously.

What should an alternative-data contract preserve?

It should make authorised uses, redistribution, retention, deletion, audit rights, service levels, correction procedures, and provenance responsibilities clear. Technical teams also need stable schemas, change notifications, and historical-version access. Legal terms and data engineering are connected: both determine whether research can be reproduced and safely operated.

How should a team retire a dataset?

Document why the dataset no longer meets legal, quality, cost, coverage, or incremental-value requirements. Identify every model, alert, report, and stored derivative that depends on it, then follow contractual retention and deletion obligations. Preserve permitted validation records and the decision history. A deliberate retirement process avoids silently replacing one feature with a superficially similar dataset whose meaning and historical behaviour are different.