Not all insider trades are created equal. A CEO spending $3 million of personal money on open-market shares is a different event than a director exercising options that were granted three years ago. Both create SEC filings. Both show up in insider trading databases. But they carry very different informational weight.

The challenge is separating the two, at scale, in real time, across every filing the SEC publishes. It matters because insider buying has historically beaten the market, but only when you filter for the right trades.

The problem: 4 million filings, 85% noise

Over the last decade, the SEC has received over 4 million Form 4 insider trading filings. These cover every transaction by directors, officers, and 10%+ shareholders of US public companies.

The vast majority are routine:

• Stock grants and awards (~35% of all filings)
• Sales, most through pre-arranged automatic selling schedules (~40%)
• Options exercises as part of compensation (~10%)

Only about 15% represent what we're looking for: discretionary open-market equity purchases. An insider choosing to put their own money in.

But even within that 15%, the range is wide. A $15,000 purchase by a director who already owns millions of shares is a very different signal than a $2 million purchase by a CEO that increases their holdings by 40%.

You can't just show everything and call it "insider signals." You need a way to rank them.

The approach: logistic regression on historical outcomes

Our scoring model is a logistic regression trained on over a decade of historical data. For every insider equity purchase in the training set, we know two things: the characteristics of the trade when it was filed, and what happened to the stock price afterward.

The model learns which combinations of characteristics have historically been associated with meaningful positive returns, and which combinations are associated with routine, market-tracking performance.

We chose logistic regression deliberately. It's interpretable, stable, and well-suited to this problem. More complex models (gradient boosting, neural networks) can achieve marginally better fit on training data, but at the cost of explainability. When we tell a user that a trade scored in the Elite tier, we want to be able to show them exactly why, not point to a black box.

What the model evaluates

The model evaluates every insider equity purchase across three compound signal categories:

Insider quality

Who is buying matters. The model considers:

• Role and title. C-suite executives (CEO, CFO, COO) have broader information access than directors or lower-level officers. A CEO's purchase carries different weight than a VP's. Our analysis of which roles produce the strongest signals quantifies this difference.
• Historical track record. Has this insider's past purchases been followed by positive returns? Insiders with a track record of well-timed purchases get higher scores on this dimension.

Market context

When the purchase happens matters. The model considers:

• Broad market conditions. Insider buying during fear and uncertainty has historically been more informative than buying during euphoria. A purchase during a market drawdown signals conviction that a purchase during a rally does not. The timing edge in insider signals is measurable and decays quickly after filing.
• Stock-specific context. Where is the stock relative to its recent trading range? What's the recent momentum? Buying after a decline versus buying at all-time highs sends a different signal.

Transaction strength

How much and how committed matters. The model considers:

• Trade size. Larger purchases represent more personal financial commitment.
• Holding increase percentage. A purchase that increases an insider's holdings by 40% is a stronger signal than one that increases holdings by 2%.
• Cluster patterns. Multiple insiders buying the same stock within a short window amplifies the signal. Three directors buying on the same day (as happened with Babcock & Wilcox in May 2026) is materially different from a single isolated purchase.

These three categories are combined into a single score. The score is then mapped to a tier.

The tier system

Every scored insider purchase is assigned to one of five tiers:

Elite. The highest-conviction signals. These represent approximately the top ~4% of all scored purchases. Trades where insider quality, market context, and transaction strength all align strongly. Elite-tier trades are rare, only 1,963 in the backtest sample spanning over a decade.

Strong. High-conviction signals that meet most but not all Elite criteria. More common (11,013 in the backtest sample) but still well above the baseline.

Positive. Above-average signals. Solid insider profiles with meaningful trade characteristics, but lacking the concentration of factors that push a trade into Strong or Elite. Sample size: 34,304.

Neutral. Baseline. These trades don't stand out positively or negatively. They represent the largest group (122,125) and perform roughly in line with the broad market.

Caution. Below-baseline signals. Trades with characteristics historically associated with weaker outcomes: often small purchases by insiders with poor track records or in unfavorable market contexts. Sample size: 61,805.

An important nuance about the top three tiers: Elite, Strong, and Positive signals show median returns ranging from 29% to 47% over a 63-day window in the historical backtest. The tiers don't represent dramatically different return magnitudes. They represent confidence and selectivity.

Elite is not "3x better than Positive." Elite is more curated, a stricter filter that accepts fewer trades, each meeting a higher threshold across all three signal categories. Think of it as increasing the purity of the signal, not the magnitude of the return.

What the model does NOT do

The model does not predict stock prices, issue recommendations, or guarantee outcomes.

• It does not predict stock prices. The model scores the historical profile of a trade, not the future direction of a stock.
• It does not issue buy or sell recommendations. A high score means "trades with similar characteristics have historically been followed by positive returns." It does not mean "buy this stock."
• It does not guarantee outcomes. Past patterns can break. Markets change. A trade that scores Elite can lose money. The model identifies historical tendencies, not certainties.

Validation

The model was trained and validated on over a decade of data covering 273,064 insider purchases with full price coverage. The validation approach:

• Out-of-sample testing. The model was not evaluated on the same data it was trained on. Training and validation sets were separated to prevent overfitting.
• Statistical significance. Elite-tier signals achieved a 69% hit rate on 30%+ returns within 3 months, versus a 17% baseline, roughly 4x the rate. This gap is statistically significant and persistent across different time windows within the dataset.
• Consistency. The model's tier rankings maintained their relative ordering across different market regimes: bull markets, bear markets, and sideways markets. The signal didn't only work in one type of environment.

For a detailed look at the full backtest results, including cumulative returns versus the S&P 500 over the decade, see our article on what happens when you invest $100 in every high-conviction signal.

Why transparency matters

Every score shows its three-factor breakdown, and every filing links to the original SEC source. Many financial products use phrases like "proprietary algorithm" or "AI-powered analysis" as marketing language without showing what's actually happening under the hood. We think that's the wrong approach.

Insider trading data is inherently transparent. It's public information filed with a government agency. A tool built on transparent data should itself be transparent.

That means:

• Every score shows its three-factor breakdown
• Every filing links to the original SEC source
• The tier system is explained, not hidden
• The backtest methodology is published with full detail
• The model's limitations are stated clearly, not buried in fine print

If you're going to use insider data as an input to your investment research, you should be able to see exactly how it's being processed and why a particular trade scored the way it did. That's the standard we hold ourselves to.

See today's highest-scoring insider filings. Start your free 30-day trial at insidersignals.io.

Related Reading

• What a Decade of SEC Insider Trading Data Actually Reveals: What we found when we processed over 4 million filings and separated the 15% signal from the 85% noise.
• Insider Buying by Sector: Where the Smart Money Is Going in 2026: Which sectors are seeing the highest concentration of Elite and Strong signals this year.
• How to Read a SEC Form 4 Filing: A step-by-step walkthrough of every field in the filing that US insiders must submit within 48 hours of a transaction.

For informational purposes only. Not investment advice.