Survivorship bias in betting is judging a tipster, a system or a method by the records still standing, while the ones that failed have dropped out of sight. A group with its losers removed always looks better than it did at the start, so a long winning record can be what luck left after the sorting.
The betting strategy hub covers the other ways a record misleads.
How survivorship bias builds a perfect record
Example: Illustrative numbers, describing a known trick rather than any real service. A newsletter sends a free tip on a two-way game to 3,200 people, half told Team A and half Team B. After each game it drops everyone who got the loser and splits the rest again, five games running.
| Game | People who have seen only winners |
|---|---|
| 1 | 3,200 / 2 = 1,600 |
| 2 | 1,600 / 2 = 800 |
| 3 | 800 / 2 = 400 |
| 4 | 400 / 2 = 200 |
| 5 | 200 / 2 = 100 |
Those 100 people have seen five straight winners from a sender with no skill, and they get the paid offer. Nothing they saw was false: they survived a filter they never saw.
Luck does the same with no plan behind it. Take an illustrative 640 tipsters with no edge, each on selections with a true 50% chance. Each goes 5 from 5 with chance 0.5 x 0.5 x 0.5 x 0.5 x 0.5 = 3.1%, so about 640 x 0.03125 = 20 will, and those 20 are the ones you hear about.
How it inflates tipster and system records
Four filters do most of the damage:
- Services that lose money tend to close, and their records go with them.
- A record restarted after a bad run, "since the new model", hides its own losing stretch.
- A leaderboard that ranks hundreds of records shows the top, and with hundreds competing, the top is mostly luck.
- A system for sale is the version that worked on past data. The versions that failed in testing were never offered.
Ask for every tip from the first day, with the time and price of each, and recount the record as how to check a tipster record sets out.
Risk: Betting involves risk. Past results are no guarantee of future results, and a record that came through a filter you cannot see proves even less. See responsible gambling for limits and support.
How to avoid survivorship bias in a backtest
A backtest breeds its own survivors whenever hindsight picks the data or the rule. Three questions catch most of it:
- Was every runner, team and bookmaker that existed on each date in the data, or only those still around today?
- Are the scratchings, voids and abandoned games still in the results, settled the way your bet would have been?
- How many versions of the rule did you try? If you tested 40 and kept the best, that rule survived your own search, so test it again on a later period it has never seen.
Then judge the rule only on results after the date you settled it. Backtesting betting strategies covers the full method, including where survivors hide in old price data.
Terms people confuse with survivorship bias
Cherry picking is choosing the good results on purpose, while survivorship needs nobody to choose: the failures drop out by themselves. A small sample is a separate problem, since a record can be complete but short, or long but filtered. And the gambler's fallacy is about what you expect after a streak, where survivorship is about which streaks you get to see. The two meet when a record that survived a hot run invites the hot hand belief that it will carry on, which the gambler's fallacy entry takes apart.