Survivorship Bias: How Sandra's 14 "Perfect" Stocks Hid 12 Losers

It's April 14, 2026. My colleague texts me.

"Dom, look at my strategy. +510% over 11 years. Check my math."

I checked. The math was clean.

Then I checked it again, differently. +28.5%.

Same strategy. Same rules. One difference: I stopped cheating.

Welcome to the most expensive cognitive bias in finance.

The Trap in One Paragraph

Sandra picked 15 US quality stocks in 2026. Netflix. Meta. Broadcom. Axon. Every one a winner. She didn't pick Sears. She didn't pick Peloton. She didn't pick SVB. Those went to zero. Of course she didn't.

Here's the problem. In 2015, nobody knew. Netflix was struggling with subscribers. Axon was a stun-gun company nobody had heard of. Broadcom was pre-AI boring. Meta was still Facebook, losing the mobile ad war.

A 2015-Sandra running the same filter on 2015 data would have picked 30-50 names. Some would have been winners. Some would have been Peloton.

Why Am I Telling You This?

Because every YouTube trading strategy you've ever seen does this. Every course. Every Twitter thread.

"I tested this on Netflix, Meta, Amazon, Nvidia, and Google and it made 800%."

Sure it did. Those stocks made 800% on their own. The strategy didn't do anything.

The Test I Ran

I took Sandra's exact execution rules. Dip at minus 20 percent. Scale up at minus 30. Deepest tranche at minus 40. Hold forever.

Then I ran two tests.

Test 1 — her universe: Apply rules to her 15 handpicked 2026 stocks. Test 2 — honest universe: Each year starting 2019, run Sandra's quality screen with only that year's data available. Pick top 10. Apply the same rules.

The actual portfolio numbers from my backtest (v2.1 Pure Hold execution):

Test	Deployed	Final	Return	CAGR
Test 1 — Sandra's 14 picks (11y)	$2.48M	$15.1M	+510%	17.5%
Test 2 — honest rolling (7y)	$2.69M	$3.45M	+28.5%	3.6%
SPY buy and hold (7y)	—	—	+212%	17.0%

Read that again. The same rules on a 2026-filtered list made +510%. The same rules on a 2019-filtered list made +28%. SPY just sitting there did 212%.

The gap between Test 1 and Test 2 is pure survivorship bias. 482 percentage points of it.

How Survivorship Bias Works in Your Head

You see a chart of a great company. You mentally write a story. "They had a moat. Management was smart. The trend was obvious."

What you don't see: the 10 companies that looked identical in 2015 and died.

In 2015, here's what you'd have said about:

WeWork: Coworking is the future. Adam Neumann is visionary. They're taking over Manhattan.
Peloton: Connected fitness is inevitable. Millennials hate gyms.
Zillow: Tech disrupts real estate. iBuying is the future.
GE: Largest industrial. Six Sigma. Dividend for decades.
Intel: Silicon is their middle name. Moore's Law.

All of those were quality names. Respected. Recommended. Peak-cycle darlings. Most are down 60-95% today.

When you pick your universe in 2026, you skip those by instinct. You just remove them from consideration.

That instinct is your survivorship bias at work.

One More Way to Feel It

Look at the S&P 500 today. 500 companies. Looks stable.

Now look at the S&P 500 from 2000. Only 167 of those companies are still in the index. The other 333 either got acquired, went bankrupt, or were replaced.

Enron, Lehman, Bear Stearns, AIG, Kodak, Circuit City. Names you knew. All gone.

But when someone shows you a "S&P 500 backtest from 2000," they usually mean the current 500 companies, backdated. Which is not the S&P 500 of 2000.

Remember: The companies that made the list today are not the companies that were on the list when your backtest started.

The Cure — Three Checks Before You Trust a Backtest

1. Ask when the universe was picked. If the answer is "today" or "recent," the backtest is contaminated.

2. Check if delisted stocks are included. Proper databases like CRSP include dead companies. Most free tools don't.

3. Run a point-in-time screen. Filter in 2015 data only what was knowable in 2015. Repeat for 2016, 2017. Never ever use 2026 information to make 2015 decisions.

For my own tests, I use SimFin which includes delisted companies, and I enforce strict publish-date cutoffs so fundamentals are only used after they were public.

The Trading Strategy Test

Want to know if a backtest has this problem? Ask one question:

"If I apply your strategy in 2015 using only 2015 data, which 10 stocks does the filter pick?"

If the answer names 2024-style winners like NVIDIA, Meta, Axon, that's a red flag. The person tuned their filter on 2026 data.

If the answer names boring 2015 names like Oracle, Visa, Costco, Johnson and Johnson, that's an honest answer. Some of those won. Some lost to inflation. That's real life.

What This Means for Sandra

I told Sandra. She took it well.

Her selection process (the fundamentals filter, the moat assessment, the management scoring) is still good. A lot of what she does is real edge. But the +510 percent number was inflated by survivorship.

Realistic forward expectation for her strategy, if she picks well each year: maybe 8-15 percent per year. Similar to SPY, maybe slightly above. Not the 17.5 percent the backtest implied.

The difference between "my strategy beats SPY by 600 percent" and "my strategy maybe matches SPY with more work" is huge. It changes whether you stop at the index or keep picking stocks.

Your Turn

Next time someone shows you a backtest, ask:

"When was the universe picked?"
"Does it include delisted companies?"
"What does the strategy pick in 2015 with 2015 data only?"

If they can't answer, the backtest is probably inflated by survivorship bias. How much? Based on my Sandra test: up to 482 percentage points over 11 years.

That is not a rounding error. That is everything.

-> Read the Pillar: The 10 Deadliest Biases — back to the main article

-> Look-Ahead Bias — Claude Knows Too Much — coming next week

-> See the real Sandra backtest tables — full numbers in the pillar