A filer with a good record in a sector that doubled has told you almost nothing yet. The decomposition that separates the three sources is standard and the arithmetic is not the hard part. The hard parts are the sample size you actually have, the fact that a large share of the names at the top of any insider board are not individuals at all, and the reference portfolio you choose, which quietly decides your answer before you fit anything.
What the board hands you and what it does not
The Top Insiders tab is a filer-level roll-up, and it is worth being precise about which fields are populated because the attribution depends entirely on that. On the capture in front of me the table carried columns for insider, titles, tickers, 30d Buy $, Holdings $, Top % Float, Trades and Win Rate. Dollars, tickers and trade counts were populated. Holdings $, Top % Float and Win Rate were blank on every visible row.
So the board gives you an identity, an instrument, a size and a count. It does not give you a return series, and the per-filer win rate that would be the crudest possible summary of one was not there to read. The module's own forward-return work lives on the Performance tab, where the attribution is run against SPY at 7, 30, 90 and 365 day horizons across the scored filing population rather than per filer. That is a different object from a track record for one person, and it is the right thing to check before assuming the platform has already done your attribution for you.
The trade counts set the ceiling on everything downstream. Across the eight visible rows they read one, six, two, three, one, twelve, two, two. That is the seven day snapshot rather than a career, but it is a useful reminder of the shape of the distribution you will be fitting on: a long tail of filers with a handful of lifetime open-market transactions and a thin head with enough observations to say anything.
The three-term decomposition and where the residual goes
The construction is the ordinary one. For each of the filer's transactions, define an event window, compute the realised return of the name over that window, and regress the series on a market factor and a sector factor. What is left is the residual, and the residual is the only thing you are entitled to call selection skill.
Three specification choices do most of the work. The first is the window. A track record measured at 30 days and the same record measured at a year are different records, and a filer who looks skilled at one horizon and flat at the other is telling you their edge is about timing rather than about the company. Run both and report both.
The second is the sector proxy. If you use a broad sector index you will attribute to sector any within-sector style tilt the filer has, typically toward small and illiquid, and the residual will absorb the size premium as though it were judgement. If you use a size-matched and sector-matched basket you strip more out, and the residual gets smaller and more honest.
The third is whether you weight by dollars. An equal-weighted record answers whether the filer picks well. A dollar-weighted record answers whether they bet big when they were right, which is the question that matters if you intend to mirror them. They can disagree completely, and the top of this board is a good illustration of why: one row was 100.00 M USD in a single transaction, another was 41.35 M USD spread across twelve.
Entity filers break the skill interpretation
Look again at the names on that board. Alongside two officers there is an investment fund, a partnership general partner, a corporate entity and a growth fund cooperative. These file because they crossed ten percent, not because they run the company, and their transactions are frequently negotiated rather than executed in the open market: primary issuance, structured participation, or the mechanical consequence of a transaction agreed months earlier.
Attributing a residual to that kind of filer produces a number, and the number means something, but it is not selection skill in the sense your investment committee will assume when they read the word. It is closer to a measure of how well the terms of a negotiated deal were set. Screening on the Titles column before you attribute is not a refinement, it is a precondition, and the population you keep should be documented as part of the method.
There is a related trap in the ticker column. Two of the eight visible rows pointed at the same underlying name. If you build a filer-level panel without deduplicating by name and date you will double-count a single situation, and since the top of the board is where the dollars concentrate, that duplication lands precisely where it does the most damage to your estimates.
What the standard errors do to the answer
This is the part that decides whether the whole exercise is worth running for a given filer. With a dozen observations, a residual mean of any plausible magnitude will have a confidence interval that spans zero comfortably. That is not a reason to skip the decomposition, it is a reason to report the interval rather than the point estimate, and to stop pretending a ranking of point estimates across filers is a ranking of anything.
Two corrections are worth building in from the start. Event windows drawn from the same filer overlap when the filer buys repeatedly in a short period, so the observations are not independent and the naive standard error is too small. And the cross-section of filers you are testing is enormous, so the best-looking residual in the population is the one you should trust least. If you are going to rank, rank on a shrunk estimate that pulls each filer toward the population mean in proportion to how little data they have, and state the shrinkage in the documentation.
Writing it down so it survives the review
A skill attribution is a claim you will be asked to defend after a position in a mirrored name goes wrong, and the defence has to have been written before the loss.
Record the filer population and the exact inclusion rule, including the officer-versus-entity screen and the minimum transaction count. Record the benchmark and sector proxy, and note that the platform's own attribution compares against SPY, so if you have chosen something else your numbers will not tie to the screen and somebody will eventually ask why. Record the event window and the convention for the entry timestamp. Record the deduplication rule for same-name transactions.
Then record the thing most people leave out, which is the set of filers you looked at and rejected. An attribution run over a population you selected after seeing the returns is not an attribution, and the only evidence that you did not do that is a contemporaneous list of everyone who was in the universe when you started. Absent that list, a strong residual is indistinguishable from a well-executed search, and the difference is the entire value of the exercise.