Back to Editorial
2026-08-03//InsiderAlpha Research//Updated 2026-08-03

Where Insider Buying Matters Most: A Sector-by-Sector Form 4 Study

A matched-company study of 2,948 mature purchase events finds the clearest buy-versus-sale separation in Technology and Healthcare—and little evidence for one universal insider signal.

Most insider-buying research begins with the wrong abstraction. It asks whether reported purchases “work” as one homogeneous signal across the market.

Companies do not operate inside one homogeneous information environment. A biotechnology board, a software executive and a regulated financial officer face different disclosure regimes, valuation uncertainty, capital cycles and information asymmetries. If reported purchases contain information, there is no reason to assume that information should travel through every sector in the same way.

We examined 2,948 sector-qualified company-day purchase events from the July 29, 2026 InsiderAlpha Form 4 Research Database. The raw sector ranking favored Technology and Healthcare. More importantly, those two sectors retained the clearest separation when purchases were compared with sales inside the same tickers.

The central result is not that every technology or healthcare purchase is bullish. It is that the difference between reported purchases and sales was materially larger in those sectors than in most of the sample.

The key finding

Among companies with enough purchase and sale history to support an internal comparison, the median 180-day purchase-minus-sale spread was:

  • Technology: +17.82 percentage points across 24 matched tickers;
  • Healthcare: +15.77 percentage points across 14 matched tickers;
  • Consumer Cyclical: +9.76 percentage points across 18 tickers;
  • Financial Services: +6.53 percentage points across 17 tickers;
  • Industrials: -3.07 percentage points across 23 tickers;
  • Energy: -4.63 percentage points across nine tickers.

Technology and Healthcare also survived an issuer-level bootstrap: their 95% intervals remained above zero. That makes the finding more credible than a ranking driven only by one unusually active company or one right-tail return.

It does not convert the result into causal evidence or an executable strategy.

What counts as a purchase

We use SEC transaction code P. Under the official Form 4 instructions, P means an open-market or private purchase of a derivative or non-derivative security. Code S similarly covers open-market or private sales.

That wording is important. A code P screen is a useful directional filter, but it should not be represented as a guarantee that every transaction occurred anonymously on a public exchange.

Awards, option exercises, gifts and tax-withholding dispositions are not included in the purchase-versus-sale comparison.

Sample and unit of analysis

The study uses the versioned July 29, 2026 database release and the following protocol:

  1. Start on January 1, 2007, when purchase coverage becomes materially denser.
  2. Require non-null sector metadata from the consistent yfinance_cache source.
  3. Keep transaction codes P and S for the matched comparison.
  4. Collapse multiple transaction lines to one ticker–transaction date–transaction code company-day.
  5. Use the median post-transaction return when several lines share the same company-day.
  6. Require a calculated 180-day outcome.
  7. For the matched comparison, require at least three mature purchase days and ten mature sale days per ticker.

The mature purchase sample contains 2,948 company-days across 232 issuers, dated from February 13, 2007 through January 16, 2026.

Why company-days rather than filing lines? One Form 4 can contain multiple securities and transaction rows. Treating each line as an independent market event would give excessive weight to complex filings and prolific reporters.

First view: outcomes after purchase days

The table below describes what happened after code P company-days. It does not subtract a market or sector benchmark.

SectorMature purchase daysIssuersMedian 30DMedian 90DMedian 180DPositive at 180D
Technology25440+2.11%+10.63%+19.90%72.8%
Healthcare20833+5.09%+11.52%+16.54%66.8%
Energy88115+0.46%+5.92%+13.58%63.2%
Consumer Cyclical16022+2.08%+5.78%+10.29%67.5%
Industrials27531+2.35%+4.96%+8.24%68.7%
Real Estate51216+1.79%+3.99%+6.50%65.4%
Financial Services41439+0.68%+3.79%+4.01%63.0%
Consumer Defensive8612-1.45%+2.77%+1.72%53.5%

Communication Services, Basic Materials and Utilities had fewer than 100 mature purchase days and should be treated as thinner descriptive cohorts.

Why the raw ranking is not enough

Absolute returns mix the insider event with the market regime in which it occurred. The Energy cohort is also highly concentrated: 881 company-days came from only 15 issuers. A sector with repeated purchases from a few companies can look more stable than its true cross-sectional breadth warrants.

For that reason, we did not use the raw ranking as the primary result. We asked a harder question: within companies that produced both purchase and sale observations, how different were the later outcomes?

Matched-company purchase-versus-sale comparison

For every eligible ticker, we calculated its median outcome after purchase days and its median outcome after sale days. We then measured:

ticker spread = median return after P − median return after S

The sector statistic is the median of those ticker-level spreads. This gives each eligible company one vote, rather than allowing the most active filers to dominate the result.

SectorMatched tickers30D spread90D spread180D spreadTickers where purchases led
Technology24+2.04 pp+9.75 pp+17.82 pp79.2%
Healthcare14+3.94 pp+15.49 pp+15.77 pp78.6%
Consumer Cyclical18+0.60 pp+3.13 pp+9.76 pp66.7%
Financial Services17+0.08 pp+2.47 pp+6.53 pp52.9%
Real Estate10+1.87 pp+3.54 pp+3.63 pp60.0%
Industrials23+3.09 pp+2.59 pp-3.07 pp34.8%
Energy9+1.75 pp+1.27 pp-4.63 pp44.4%
Consumer Defensive9+0.41 pp+2.89 pp-5.10 pp22.2%

The cross-horizon pattern matters. Technology and Healthcare did not rely on a single 180-day endpoint; both showed positive separation at 30, 90 and 180 days. Industrials and Energy were positive at shorter horizons but negative at 180 days in the matched-ticker median.

Robustness: resampling issuers, not rows

Transaction rows from the same issuer are correlated. A conventional row-level confidence interval would understate uncertainty.

We therefore resampled the matched tickers 10,000 times with replacement and recalculated the median 180-day spread. The random seed was fixed for reproducibility.

SectorMedian matched spreadIssuer bootstrap 95% interval
Technology+17.82 pp+7.41 to +30.52 pp
Healthcare+15.77 pp+2.25 to +21.57 pp
Consumer Cyclical+9.76 pp-1.94 to +13.12 pp
Financial Services+6.53 pp-9.26 to +18.90 pp

Only Technology and Healthcare remained entirely above zero in this diagnostic. Their intervals are wide and overlap each other, so the study cannot establish that Technology is definitively stronger than Healthcare. It can say that both separated more consistently than the other adequately represented sectors in this release.

Interpretation

Several mechanisms could produce sector differences without insiders possessing a magical forecasting ability.

Information environments differ

Healthcare and Technology often contain firms whose value depends on product milestones, intellectual property, adoption curves or research outcomes. Those environments can create wider gaps between internal operating knowledge and outside consensus.

That explanation is plausible, not proven by this dataset. The Form 4 records do not contain the insider's motive or private information set.

Purchases and sales have asymmetric motivations

Sales can fund taxes, diversification, liquidity or pre-arranged plans. Purchases generally require the insider to increase exposure, but code P can still include private transactions and does not reveal personal wealth. The matched spread may therefore reflect noisier sale motives as much as stronger purchase information.

Sector composition can masquerade as signal

Smaller companies, distressed issuers and firms with limited analyst coverage may be unevenly distributed across sectors. Without market-cap, valuation and liquidity controls, part of the observed spread may represent those characteristics rather than sector itself.

What a production model should do

The evidence argues against a universal insider_bought = 1 feature. A research model should allow the purchase effect to interact with sector and other documented context:

transaction code × sector × cluster size × owner role × ownership change

An executable model would also need external data for:

  • sector and broad-market benchmark returns;
  • market capitalization and liquidity;
  • earnings and corporate-event timing;
  • entry after the Form 4 became public;
  • delisted securities and point-in-time issuer classifications;
  • transaction costs and portfolio construction.

Those controls are not optional if the objective is to claim alpha rather than describe historical outcomes.

Limitations

This study is intentionally narrow.

  • market_return_* measures absolute security-price change from the transaction-date baseline. It is not abnormal return.
  • The matched design controls partially for issuer identity, not for market regime or time-varying firm characteristics.
  • Sector metadata comes from a selected consistent source and covers only a subset of the complete purchase population.
  • Current metadata availability can create selection or survivorship effects.
  • Repeated company-days are not independent, even after line-level deduplication.
  • A transaction-date event study cannot be interpreted as a public-information trading strategy.
  • Historical association does not establish causality or predict future returns.

The full field definitions, missingness contract and return methodology are available in the documentation.

Conclusion

Sector changed the interpretation of reported insider activity. In this metadata-qualified sample, Technology and Healthcare showed the most persistent purchase-versus-sale separation, both across horizons and after equalizing the comparison at the ticker level.

The result is useful as feature-engineering evidence, not as a list of sectors to buy blindly. It says that insider activity should be modeled conditionally—and that a database capable of preserving transaction code, sector provenance, reporting-owner identity and post-event outcomes can test that condition directly.

The complete versioned dataset is available in CSV, Parquet and DuckDB formats through InsiderAlpha Research Data. For targeted issuer, owner and transaction queries, see the API access options.

Tags
Insider BuyingSector AnalysisForm 4TechnologyHealthcareEvent Study