What we tested

No experiments are recorded yet.

Earlier studies

Run before the experiment registry existed. Source: the experiment ledger.

• Feroldi / Piotroski quality gates as extra filters (Feroldi gates)

  FAIL

  Eight gate settings, 2012-2021, in-sample. Every gate raised downside capture (baseline 0.68; gates 0.74 to 0.91), the opposite of the goal. None earned a place.

• ROCE as the tiebreaker instead of ROE (QUA-190 A)

  FAIL

  Identical results in all six folds. Gross profit never ties in practice, so the tiebreaker never ran. It changed nothing.

• Composite rank: 0.7 × gross profit + 0.3 × ROCE (QUA-190 B)

  FAIL

  The book changed (about 6.7 of 30 names per rebalance), but downside capture rose from 0.71 to 0.79 (worse) and Sharpe fell slightly.

How we test

Walk-forward out-of-sample

We split history into consecutive folds. Each fold is judged only on data after the period used to choose the rule, so a rule cannot be tuned on the answer.

Deflated Sharpe

The more ideas we try, the more likely one looks good by luck. The Deflated Sharpe ratio lowers the score for the number of trials, so a lucky result does not pass.

Cost model

Every trade pays an assumed trading cost. All scores are net of cost. We never judge an idea on a zero-cost number.

Point-in-time data

A company fact counts only from the day it was filed with the SEC. Later restated values are not used, so a test cannot see the future.