The email we didn't want to receive
A subscriber sent us an email with a chart attached. They had rebuilt our Global Growth Cycle Enhanced Momentum backtest from scratch, and their curve matched our published version almost exactly. Then they made one change: instead of the OECD Composite Leading Indicator data as it looks today, they plugged in the numbers as they were originally published. The strategy fell apart. They asked, politely, whether we were confident in our data.
We were not.
What the OECD's CLI does after every release
GGCEM uses the OECD Composite Leading Indicator to set its risk-on or risk-off posture, drawing diffusion signals across 17 countries. The catch is what happens after each monthly release: the OECD refits its trend model on the full available sample and reissues the entire history. A first print and its final revised value can tell different stories, and turning points quietly move to where they should have been.
Our backtest handled timing correctly. It never read a number before its release date. But it read today's revised values for every historical month. That's information nobody held at the time.
What we found when we checked
We rebuilt the test ourselves rather than take their chart on faith. ALFRED, the St. Louis Fed's public real-time archive, keeps actual publication snapshots of the CLI series starting mid-2018. We replayed GGCEM twice on our production engine: once with data as it looks today, once with data exactly as it was published. Same strategy, same engine, same costs.
Since 2018, the risk regime call differs between the two data sets in 19% of months. On revised data, the strategy shows a Sharpe of 1.19 and a worst drawdown of -11%. On as-published data: Sharpe 0.82, worst drawdown -34%. That's roughly the risk profile of just holding the index.
The detail that stings: the 2020 crash sidestep, the most impressive moment on the published curve, exists only in the revised data. The first prints said risk-on straight through the drawdown, so a real-time follower rode the whole thing down. The 2022 bond defense is a different story. That signal was identical across both data sets, which means that part of the record was genuine in real time.
The signal isn't worthless. It's just much weaker than the published curve implies.
An old trap with a documented history
Economists have known about this problem since at least 1991, when Diebold and Rudebusch demonstrated that the composite leading index's real-time forecasting record collapses compared to its revised-data record. Their chapter PDF is publicly available.
AllocateSmartly, which tracks the original Global Growth Cycle strategy, uses earliest-vintage economic data on their platform. After testing it on vintage CLI data, they found the strategy was "still effective, just not as effective" as the author's original test suggested. Our version sat above even that, because we were working with fully revised data.
The uncomfortable part is that live signals were never wrong. They always used first publications, because that's all that exists in the moment. The look-ahead lived only in the backtest history. The published curve was quietly promising something the live strategy was never going to deliver.
What changed on the site
We've published a data caveat and a real-time data check on the GGCEM page, with both curves and numbers visible side by side. GGCEM is out of our Top Performers rankings while we decide its future. The strategy description now states plainly that the backtest uses revised macro data.
History before 2018 cannot be checked this way. No public archive of those earlier vintages exists, so the deeper backtest keeps its disclosure instead of a correction.
The practical lesson
If a strategy's edge comes from macro data, ask which vintage the backtest used. Price data doesn't get revised. Economic data does, and the revisions flatter every backtest built on them.
We'd rather publish a weaker true number than a stronger false one. Thanks to the subscriber who forced the question.
Past performance does not guarantee future results.