Isolating the Insider Signal: Why Most Form 4 Buys Contain Limited Predictive Information
A quantitative breakdown of how to filter 1.5 million SEC filings to separate routine compliance trading from genuine, alpha-generating conviction.
A common trap for novice quantitative analysts is treating all SEC Form 4 filings as equal data points. The naive assumption is that if an insider buys shares, it acts as a bullish signal for the underlying equity.
In reality, raw insider data is not a signal. It is a feature reservoir.
Routine purchases, minor portfolio rebalancing, automated 10b5-1 plans, and trivial dollar amounts create a dataset where the vast majority of transactions exhibit near-zero predictive correlation with future returns.
To demonstrate how to construct a robust quantitative feature, our research team analyzed over 1.5 million historical Form 4 records from the InsiderAlpha Master Dataset. We applied sequential filters to strip away the statistical noise and isolate the true Insider Conviction Signal.
The Baseline: The Raw Dataset
Before applying any filters, we ran a baseline backtest across all Open Market Purchases (Transaction Code = P) in our universe (US Equities > $50M Market Cap, excluding biotechs).
Baseline Metrics (Unfiltered):
- Total Events (N): 1,245,892
- Median 90-Day Excess Return: +0.15% (vs. Russell 2000)
- Win Rate: 50.4%
As expected, buying every Form 4 filing blindly yields returns indistinguishable from a coin flip. The transaction costs alone would render the strategy deeply unprofitable.
Filter 1: Dollar Conviction Buckets
The first and most crucial filter is Dollar Conviction, defined as (Total Purchase Value / Company Market Cap).
Corporate governance policies frequently require executives to hold a minimum amount of equity. A VP buying $15,000 of stock in a $5 Billion company is likely satisfying a board mandate, not making a conviction bet. To find the optimal threshold without data mining, we evaluated the returns across predefined conviction buckets:
| Conviction Bracket | Median 90-Day Excess | Win Rate |
|---|---|---|
| < 0.01% | +0.2% | 50.8% |
| 0.01% - 0.05% | +1.8% | 51.5% |
| 0.05% - 0.10% | +5.2% | 54.1% |
| > 0.10% | +8.4% | 56.6% |
The data clearly shows an inflection point at 0.05%. We applied this as our base filter, eliminating over 91% of the baseline transactions.
Filter 2: Insider Quality
Not all insiders possess the same predictive track record. As detailed in our previous CEO vs CFO analysis, CFO purchases historically demonstrate a more measurable performance advantage compared to CEOs (who may buy for PR signaling) or lower-level Directors.
Instead of treating all insiders equally, executive roles were weighted according to their historical predictive performance, prioritizing CFOs and specialized Directors.
Filter 3: Synchronization (The Cluster)
A single executive making a high-conviction buy is a strong signal. Multiple executives executing this exact pattern simultaneously is an anomaly. We applied our synchronization filter, requiring at least three distinct insiders to purchase shares within a 72-hour window.
Filter 4: Market Context (Value vs. Falling Knife)
Insiders are notorious value investors. However, buying a falling knife is dangerous. To avoid the overfitting and data-mining critiques common in retail backtests, we established an a priori hypothesis tested on an out-of-sample data segment: High-conviction insider buying is most effective when downside momentum is exhausted.
We applied a basic technical filter: The stock must be trading below its 200-day moving average, but with a positive 5-day Rate of Change to ensure price stabilization at the time of the filing.
The Final Combined Signal
When we combine these four features (Conviction > 0.05%, CFO/Director Weighting, Cluster Size >= 3, and Price Stabilization), we intentionally sacrifice sample size for much higher signal purity to prevent curve-fitting:
- Independent Corporate Events (N): 512
- Median 90-Day Excess Return: +12.40%
- Win Rate (90D): 63.5%
- Maximum Drawdown (Portfolio Level): -14.2%
(Note: The Portfolio Drawdown assumes an equal-weighted portfolio of signaled equities, held for a strict 90-day period with no stop-losses applied).
The Waterfall of Alpha
The impact of each filter on the dataset becomes strikingly clear when viewed sequentially:
| Model Progression | Events (N) | Median 90D Excess | Win Rate |
|---|---|---|---|
| All Form 4s (Baseline) | 1,245,892 | +0.15% | 50.4% |
| + Dollar Conviction | 112,430 | +5.20% | 54.1% |
| + Cluster Synchronization | 3,421 | +9.80% | 58.7% |
| + Final Score (Market Context) | 512 | +12.40% | 63.5% |
Conclusion: Feature Engineering Over Raw Data
The progression of the data reveals a fundamental truth about alternative data modeling:
Raw SEC Filing + Role Weight + Dollar Conviction + Cluster + Market Context = InsiderAlpha Score
By applying logical, non-correlated filters, we transformed a baseline strategy with zero historical performance into a model historically generating a +12.40% median excess return.
If you are feeding raw Form 4 data into a machine learning model or a quantitative backtest, your results will be heavily degraded by statistical noise.
To build sophisticated models without spending months cleaning SEC Edgar data, acquire the InsiderAlpha Master Dataset. It provides pre-calculated Dollar Conviction, Cluster Sizes, and Executive Roles for 1.5 million transactions, allowing you to focus on alpha generation rather than data wrangling.