Our backtest says 20% a year. The honest number is 13% — and it loses to the S&P 500. Here's why we tell you that.

Pillar: Methodology / how backtests lie · Educational, not financial advice

Every quant tool can show you a backtest where it beats the market. That's not a brag. It's almost a tautology — backtests are easy to win, because the most common way to run one quietly cheats. This post is about the most expensive of those cheats, using our own numbers as the example.

Two numbers from the same strategy

Run our best configuration one way, and it returns about 20.5% a year and beats the S&P 500 by roughly +3.5 percentage points.

Run the same strategy honestly, and it returns about 13.3% a year and loses to the S&P 500 by roughly −3.8 percentage points.

Same rules. Same stocks. Same code. A ~7-point-a-year swing between "market-beating" and "market-lagging." The gap is not noise. It has a name.

The cheat is called survivorship bias

The flattering version of a backtest lets the model pick from today's list of companies — the S&P 500 as it exists now. The honest version forces the model to pick only from the companies that actually existed on each historical date, with no knowledge of who would later thrive, get acquired, or go bankrupt.

That difference is everything. Today's index is a list of winners — the companies that survived. Letting a backtest shop from that list is like grading a stock-picker who was handed the answer key: of course the "1998 portfolio" looks brilliant if it's only allowed to buy the companies we now know made it to 2026.

The honest run rebuilds the investable universe as it stood on each day in the past — survivors and casualties alike — and makes the model choose blind. That ~7-point gap is the measurement of the bias. It's how much of the flattering result was the answer key, not the strategy.

Why we publish the worse number

Because the flattering number would be a lie, and the whole point of this tool is to not lie to you.

It would be trivial to put "beats the S&P by 3.5% a year" on the homepage. The static backtest genuinely produces it. But we know it's survivorship-inflated, so showing it would be selling you a result we don't believe. The honest figure — point-in-time, survivors-and-all, no answer key — is that the strategy currently trails the index. That's the number we stand behind, and it's the one that tells you what the tool is actually for.

So what is it for, if it doesn't beat the market?

This is the part worth sitting with. Divergia is not a "beat the market" engine, and we won't pretend it is. It's a discipline engine: a conservative, transparent way to ask "what is this business plausibly worth, on honest data, and should I wait for a better price?" — and to get an answer that includes "no" and "I don't know."

The value isn't a return you can't verify. It's that every number it gives you survived the model trying to not flatter itself: honest data, conservative assumptions, and a backtest that refuses to peek at the answer key.

A tool that beats the market in its own backtest is the default. A tool that tells you its honest backtest loses — and explains exactly why the other number was inflated — is the rare one. We'd rather be the rare one.

Educational/informational content generated by a quantitative model — not financial advice, not a recommendation to buy or sell, and not personalised to your situation. Backtested results are hypothetical, do not represent actual trading, and are not a guarantee of future results. Figures are model outputs under stated assumptions. Do your own research.

See the honest read on any ticker — verdicts included "wait" and "I don't know": https://divergia.ai/