We Published Wrong Backtest Numbers. Here's the Correction.
We Published Wrong Backtest Numbers. Here's the Correction.
On September 20, 2026, we published a QQQ golden-cross backtest showing +273.2%. On September 23, we replaced it with +238.5% and explained why. This post is the full account — what broke, how we found it, and what changed in our process.
What the wrong number was
- Published 2026-09-20: 4 trades, +273.2% total return, 21.0% CAGR, −28.6% max drawdown, 81.8% time in market.
- Corrected 2026-09-23: 3 trades, +238.5% total return, 19.3% CAGR, −22.8% max drawdown, 74.0% time in market.
Same strategy, same ticker, same window. Different answer — because the first run had two bugs.
Bug 1: a phantom first-bar signal
The engine evaluated a golden cross on the very first bar where both SMAs could be computed (2019-10-16). But a cross needs a transition — short below long on the previous bar, short above long on the current one. On the first computable bar there is no previous bar, so the "signal" was an artifact of the data starting, not of the market. It manufactured a fourth trade out of thin air.
Bug 2: same-bar look-ahead
The v0.1.0 loop credited the entry bar's own return to new positions (look-ahead — inflates results) and dropped the exit bar's return even though the position was still exposed that day (deflates results). On this dataset the dropped exit-day returns outweighed the phantom entry-day gains — so the corrected run came out higher than the flawed recheck (+238.5% vs +228.5%).
That surprised us, and it's why the lesson we recorded is precise: never describe a biased run as "overstated." It's unreliable. Rerun it.
How we know the new number is right
- Two independent implementations — our
stratiq-backtestengine and a standalone script — read the same 1,939-bar Yahoo Finance dataset and produced bit-identical results across all 10 headline metrics. - Every cross date verified by hand-computed SMAs at full precision.
- Year-end closes cross-checked against Macrotrends within 0.11%.
What changed in our process
The execution rules we now publish came directly out of this incident: no look-ahead (positions earn from bar i+1), no phantom first-bar signals (crosses need a genuine transition), exit-day returns always counted, and every headline backtest computed twice.
Why publish this at all?
Because the incentives in this industry run the other way: quietly fix the spreadsheet, keep the flattering number. We'd rather be the shop that corrects itself in public. If you can't trust our numbers when they're embarrassing, you shouldn't trust them when they're flattering either.
The corrected backtest, in full: QQQ Golden Cross: +238.5% vs +289.1% Buy-and-Hold.
Not financial advice. Past performance doesn't predict future results.