Evidence

Realised months, not a backtest.

A backtest replays a rule over history. What follows is what each engine actually ran, month by month, read from its performance workbook — with any month the source could not produce left as a gap.

A second lens, on the same universe: grouping companies by profile instead of by sector

Two engines, two benchmarks, two methodologies — shown one at a time, never combined.

Short engine · realised selection alpha

+4.8%

Compounded selection alpha against the S&P 500 Equal Weight / SPW Index, over 36 observed months (Aug 2023 – Jul 2026).

Trailing 12 months −8.1%, Latest month (Jul 2026) −1.9%

Compounded, never summed, from the months the engine actually ran. The full-period figure and the recent readings can differ — the path below shows how it got here.

Cumulative selection alpha, with each month beneath

Compounded pathFavourable monthUnfavourable monthUp = the shorts lagged the benchmark.

Months observed: 36Aug 2023 – Jul 2026
Favourable months: 21 of 3658.3%
Calendar gaps: NoneEvery month present in the source

Go deeper

Same months, other angles

How the months are distributedHistogram of the 36 monthly readings, from worst to best.

What each leg didThe shorted basket's monthly return against the benchmark's — the two things being differenced.

Name by nameOne dot per shorted name per month: the basket figure is an average of names that did not all move together.

Illustrative 100/90 overlay — a simulation, not alphaKept out of the headline on purpose: it blends the benchmark's own market return back in.

How to read thisSign convention, gaps, and what the source could not supply.

Sign convention: The source workbook stores short basket − benchmark, where a negative number is the favourable outcome. Every figure here uses benchmark − short basket instead, so a favourable month reads positive. The workbook’s own value is never shown.
Calendar coverage: Every month from Aug 2023 to Jul 2026 is present. The gap fields are checked on every load, so this note changes if a future delivery reopens one. The security-level file covers every month of the monthly file.
Blank tickers: 1075 of 1804 security-month rows carry a blank ticker. Names are identified by security, which is populated and unique per date throughout.
Fields left out: The security file also carries first-week and last-week performance columns. They are not consistently populated across the history, so every figure here uses the one-month performance and benchmark return columns only.
Overlay weighting: The source workbook’s chart title states a different short weighting than its formulas compute. The formulas, and this page, use the 90% figure; the title’s figure is not used anywhere.

Walk through the evidence with us

We can take you through how each engine is measured, what its months show, and what they deliberately do not.

Request a walkthrough