Filing date versus trade date
Many databases are built around filing dates because those are easy to collect. For this signal, that is not good enough. The event-time variable must be anchored to trade date, while any investable implementation must consider filing date because that is when the market learns about the trade. These are different questions:
- Trade-date analysis asks whether insiders possess useful information.
- Filing-date analysis asks whether outside investors can capture it.
A proper article should ideally show both. The trade-date version is academically cleaner. The filing-date version is what readers can actually trade, subject to slippage and liquidity.
Earnings date endogeneity
Insiders may know that earnings will be brought forward, delayed, or accompanied by guidance changes. If the public calendar is stale, the measured days_to_earnings can be wrong in a way that is correlated with the insider’s information. That is a subtle source of bias. It will not ruin every test, but it should make one modest.
Small-cap concentration
Insider signal strength is often stronger in small and less-followed firms. Those same firms may have more variable reporting practices and less rigid internal trading controls. If the 14-day premium is really just a small-cap premium wearing a lanyard, sector and size controls should expose that. At minimum, stratify results by market capitalisation.
A research design worth publishing
Baseline specification
If the data were fetched, the baseline table I would want is straightforward:
- Sample of discretionary open-market buys.
- Sort by
days_to_earnings buckets.
- Compute average and median T+90 excess return.
- Report counts, hit rates, and value-weighted notional.
- Control for size, sector, country, insider role, and filing lag.
A regression form could look like:
[
ExcessReturn_{i,t+90} = \alpha + \beta_1 NearEarnings_{i} + \beta_2 Size + \beta_3 Momentum + \beta_4 Sector + \beta_5 Country + \beta_6 InsiderRole + \epsilon
]
where NearEarnings is an indicator for the 8 to 14 day bucket, or a spline over event-time distance. One can also test non-linearity. It would not be surprising if the relationship were hump-shaped rather than monotonic. Extremely close trades, say 0 to 3 days, may be rare and distorted by restrictions. Moderately close trades, say 8 to 14 days, may be the sweet spot. Trades far from earnings may simply be less tied to a concrete catalyst.
Interaction terms that actually matter
Three interactions are especially worth testing.
Insider role × earnings proximity
CEOs and CFOs are more informed, but also more restricted and more scrutinised. Non-executive directors may trade with less direct operational knowledge but perhaps more freedom in some firms. If the premium is concentrated among one role, that tells you something useful about mechanism.
Buy size × earnings proximity
Large discretionary purchases, especially relative to the insider’s prior holdings or annual cash compensation, tend to be more informative than token buys. A 14-day premium that appears only in large buys is more credible than one driven by symbolic purchases.
Prior stock performance × earnings proximity
Contrarian insider buys after sharp drawdowns often predict rebounds. If the 14-day effect survives after controlling for prior six-month return, it is less likely to be a simple “insiders buy weakness” story.
Event-study sanity checks
Before getting lost in regressions, run a plain event study. Plot cumulative abnormal returns from filing date to T+90 for each event-time bucket. If the curves are indistinguishable until day 40 and then one suddenly levitates, your benchmark or sample construction probably needs a stern conversation.
A second useful plot is trade frequency by days_to_earnings. In a strict MAR environment, activity should collapse in the final 30 days for covered insiders. If it does not, either the sample includes many non-covered persons, the transaction coding is broad, or the earnings calendar is wrong. None of those are fatal, but each changes the interpretation.
What we can say now, before the numbers arrive
The prior from the literature
The literature gives a cautious prior in favour of conditional insider-buy alpha, but not a free pass for this exact variant. Broadly, insider purchases outperform on average, especially in smaller firms and when trades are clustered or opportunistic. There is also a large literature on post-earnings announcement drift, showing that the market does not fully incorporate earnings news immediately. Those two facts make the “pre-earnings insider buy” hypothesis plausible. They do not prove that 14 days is the magic number.
The closest intellectual neighbour is not simply insider trading research, but the literature on attention, disclosure timing, and underreaction. A pre-earnings buy could be informative because the insider anticipates a positive earnings surprise, because they expect guidance to improve, or because they know the market’s current expectations are stale. T+90 is where those channels tend to reveal themselves.
What would count as convincing evidence
Given the legal and selection issues, I would regard the following as convincing:
- a positive and statistically reliable T+90 excess return for the 8 to 14 day bucket,
- persistence after controlling for size, sector, country, insider role, and prior momentum,
- stronger effects for larger discretionary buys,
- similar direction on filing-date implementation, even if attenuated,
- no obvious concentration in a handful of illiquid microcaps,
- and a sensible cross-market pattern consistent with local dealing rules.
What would not convince me is a single pooled average with no controls, no role split, and no discussion of blackout windows. That is not a result. It is a request to be fooled.
What might falsify the thesis
Three outcomes would seriously weaken the idea.
First, if the premium disappears once trades under plans, option-related transactions, and tiny notionals are removed. Second, if the effect is entirely explained by small-cap exposure or by buying after drawdowns. Third, if filing-date returns are flat, implying that any information advantage exists only on trade date and is not monetisable by outside investors.
A fourth possibility is more interesting than fatal: the premium may be positive in the US but absent or reversed in France and the EU because MAR-style restrictions remove the cleanest opportunities. That would not kill the general concept. It would simply mean the signal is regime-specific, which is often how markets remind us that law is part of data-generating process.