Why not equal weights?
Equal weighting is attractive because it looks honest. It is also a confession that you do not know which features matter more.
That can be acceptable at the prototype stage. It is less acceptable once you have enough history to test whether one family of variables consistently improves discrimination. Equal weights also create a subtle problem: they over-reward buckets that contain many correlated sub-features. A bucket with ten mildly related inputs can dominate a bucket with three stronger but cleaner inputs unless you explicitly control its contribution.
So we weight at the bucket level first, then normalise within buckets. This keeps the score interpretable and prevents feature-count inflation from masquerading as insight.
Bucket one, signal quality at 30%
Signal quality asks a rude but necessary question: is this filing likely to reflect a discretionary information-bearing action, or is it administrative wallpaper?
This bucket includes things like:
- open-market purchase versus sale
- discretionary trade versus automatic or plan-based transaction
- direct ownership versus indirect vehicle
- one-off filing versus repetitive pattern
- clean economic exposure versus derivatives or structured instruments
- whether the filing appears complete, timely, and internally consistent
Why 30%, not 50%
Signal quality is foundational, but not all foundational variables deserve dominant weight. Its job is partly gating rather than forecasting. A poor-quality signal should be penalised heavily, yes. But once you have established that a filing is reasonably clean and discretionary, the marginal value of further quality distinctions falls.
That is why we stop at 30%. If you push this bucket too high, the model starts behaving like a filing-classification engine rather than an alpha-ranking system. It gets very good at telling you which transactions look proper, and less good at telling you which proper transactions matter most.
What tends to survive ablation
In most insider datasets, the strongest contributors inside this bucket are not exotic:
- purchase versus sale
- open-market versus non-open-market
- discretionary versus automatic
- transaction type exclusions, especially awards, tax-withholding disposals, and option exercises with immediate sale components
This is consistent with the literature. Open-market purchases have historically carried stronger information content than many categories of sales, in part because sales are motivated by diversification, tax, liquidity, and estate planning, while purchases require fresh capital and usually a clearer positive view. Seyhun’s long body of work is still the unavoidable starting point here, even if one should resist treating old US findings as eternal law.
Why this bucket is not larger
The main reason is saturation. Once obvious low-information filings are filtered or penalised, the incremental gain from ever finer classifications diminishes. You can spend months perfecting a taxonomy of derivative transactions and still learn less than you would from a simple measure of whether the trade was unusually large relative to the insider’s prior stake.
That is the first principle behind the 30/35/25/10 split: cleanliness matters, but context carries more of the residual signal.
Bucket two, transaction context at 35%
This is the largest bucket because it answers the question investors actually care about: how meaningful is the trade in context?
A €50,000 purchase can be trivial for one insider and substantial for another. Ten thousand shares can be symbolic in a mega-cap and highly consequential in a small-cap. A sale after a ten-year run-up is not the same as a sale into distress. Context converts a filing from a legal event into an economic event.
What sits inside transaction context
Typical sub-features include:
- trade size relative to prior holdings
- trade size relative to annual compensation, where available
- trade size relative to average daily volume or free float
- clustering, multiple insiders buying within a short window
- first purchase after a long period of inactivity
- repeat buying versus one-off signalling
- buying after drawdowns versus buying near highs
- concentration, whether the insider is increasing or reducing exposure meaningfully
These are not cosmetic details. They are often where persistence lives.
Why this gets the largest weight
Because context is where insider intent becomes legible.
A chief financial officer buying a modest amount every quarter under a regular pattern is less interesting than the same officer making a first meaningful open-market purchase after eighteen months of silence. Likewise, a non-executive director buying a token amount may be less informative than a divisional executive making a purchase equal to a large fraction of annual cash pay.
An ablation study tends to show this bucket doing heavy lifting in two ways:
- Cross-sectional discrimination. It helps separate merely positive filings from genuinely unusual ones.
- Rank ordering among good signals. Once poor-quality filings are down-weighted, context often determines which names rise to the top decile.
This is why transaction context gets 35%, the largest single share.
The practical edge of clustering
One contextual feature deserves special mention: cluster buying. Academic work has repeatedly found that purchases by multiple insiders over a relatively short period contain more information than isolated trades. The intuition is plain enough. Independent insiders all have reasons not to buy. When several do so anyway, coincidence becomes a less satisfying explanation.
That does not mean every cluster is gold. Small boards can create accidental clusters, and blackout windows can mechanically compress activity. But as a contextual feature, clustering has been persistent enough to earn material weight.
