Ahmer Nadeem Khan
Return to project

Backtesting design and decisions

To be written. Note, trade timing: the decision is formed after the close on day t using close-t information, and executed at the open on day t+1. To investigate: the overnight gap haircut on a signal measured at the close, and whether share count or dollar notional is held fixed between decision and fill. Also to cover: rounding of the regulatory fees, which brokers apply to the cent per trade. Also to state: we will not model recall risk, where a lender calls the shares back and forces an exit we did not choose, nor situations where a locate for a borrow cannot be obtained at all. Committed capital: the account is fully funded with $100k of initial gross capital, so no overnight debit balance arises and margin interest does not enter the model. Also to cover: spin-offs occurring inside a holding period. A long leg receives shares in the separated company that were never part of the pair and have to be liquidated, while a short leg owes those shares to the lender and must buy them in the open market, which is a forced purchase in a newly listed and usually thin security. These events are announced months in advance and when-issued trading typically begins before the ex-date, so screening on the announcement date is available to the strategy and is not lookahead, provided the announcement date rather than the completion date is the one used. Separately, parent and spun-off child are frequently both index members (ABT and ABBV, DOW and DD, RTX with CARR and OTIS, GE with GEHC and GEV, and others), and any formation window straddling the separation has the parent's history standing in for both legs, which produces cointegration by construction. This belongs with the existing decision on same-issuer and dual-class pairs. Also to cover: the ex-date convention for large distributions. Where a distribution is worth roughly 25% or more of the share price, exchange rules place the ex-date on the first business day after the payment date rather than before the record date, with due bills attaching to trades in between. AIV's $8.20 special dividend against a $32.12 close sits right on that threshold. The effective date therefore has to be read from the event record rather than inferred from the usual convention, since a seam placed on the wrong day produces a phantom return the size of the whole distribution.

One eligibility rule is already settled. A pair is rejected where one price series contains or continues the other's history, because the two legs are then partly the same data and will cointegrate by construction rather than by any economic relationship. This covers the legacy identities in the store, where DD and DD_2017 are the same company either side of the DowDuPont transaction, and it covers a formation window that straddles a separation, where the parent's history stands in for both the parent and the child. It is not a ban on parent and child as a pair once both have independent history after the split. Dual-class listings are the related case and are handled separately: they are genuinely distinct securities, so they are reported as their own predeclared stratum rather than excluded. A good control is dual classes; here they are separate tradeable securities inextricably linked to each other, so we expect strong to complete cointegration, yet no divergence for us to bet the reversion on.