A Five-Game Win Streak and the Small-Sample Trap: When the Standings Tell the Wrong Story

Over their last five games, a mid-table team suddenly won four, averaged more than 115 points per game, and was instantly branded by the media as 'the phenomenon of the season.' I didn't watch the highlights. I reopened every possession in my log file. And the truth sat in a very different number: across those five games, their opponents shot just 29.4% from three — nearly seven percentage points below those same players' career averages. Their true margin, their net rating, was only +1.6, the level of a mid-tier team, not a title contender. What is frightening is not that team, but the way a story gets built from a sample far too small to say anything. Numbers don't lie, but they don't tell stories either — and when the sample is too small, the numbers haven't even started speaking. I work as a data consultant for basketball teams. Over more than 13 years, I've learned something no classroom teaches: most conclusions in this sport are born from samples that are too small, then defended with confidence instead of evidence. The regular season is the perfect environment for that kind of mistake. There are three or four games a week, each one an emotional storm, and fans — along with part of the media — need a story to attach to a number. A player goes 5-for-7 from beyond the arc across two games and instantly becomes a 'rising sharpshooter.' A team loses three straight and instantly becomes a 'locker-room crisis.' I understand why it happens. I just don't agree with presenting it as fact. My method is simple to the point of being uncomfortable. Before concluding anything about a team, I ask: how many possessions do I actually have? An NBA game is roughly 100 possessions per team. A five-game stretch is about 500. That sounds like a lot, but once you isolate the metrics that depend on luck — opponent three-point percentage, free-throw rate in close games, or win rate in games decided by fewer than five points — 500 possessions still sits inside the noise. I don't say this to sound clever. I say it because this is the line between analysis and guesswork. There are three categories of numbers that the media routinely mistakes for form, when they are really just variance. First, opponent three-point percentage. When a team wins several in a row, people praise the defense. But opponent three-point percentage correlates very weakly with actual defensive quality over a short sample — it is driven mostly by whether opponents made or missed, not by whether the team defended well. If opponents shoot under 30% from three for five straight games, it is most likely luck, and it will revert to the mean. This is the law of regression to the mean that anyone reading data seriously has to accept. Second, win rate in close games. A team that goes 8-2 in games decided by fewer than five points is usually celebrated for 'grit.' Not so fast. Across ten such games, the total number of decisive possessions is only a few dozen. At that scale, one lucky shot or one referee's whistle is enough to flip the entire result. Long-run NBA studies show that a team's clutch win rate tends to regress hard toward 50% across seasons. In other words, 'grit' is mostly variance, not skill. Third, an individual's output. A guard shooting 45% from three over the first ten games looks spectacular. But if he only takes 4.5 attempts a game, ten games is just 45 shots. With 45 shots, the confidence interval for his true rate is so wide that you cannot distinguish a 38% shooter from a 42% one. I once pointed this out while working with a team that was considering a big contract for a player purely on the strength of a short hot streak — and the full dataset showed that rate was unsustainable. So where is the threshold? There is no magic number, but there is a principle. For stable metrics like rebound margin or assist rate, a few hundred possessions already say a lot. For probability-driven metrics like three-point shooting, you need a much bigger sample — usually a full season. And for clutch metrics, even one season is sometimes not enough. That difference is exactly my point: it is all '500 possessions,' yet the reliability changes completely depending on the nature of the metric. People look at the box score to remember a game. I look at shot quality to understand how the game would have unfolded if probability had dealt its cards more fairly. At this point someone usually objects: 'So isn't data useless, if everything can be luck?' No. Correlation is not causation — but a lack of data is not evidence of randomness either. Those are two symmetrical mistakes, and I make neither. The first mistake is reading causation into variance: seeing a win and concluding the system is good. The second is seeing a small sample, hastily calling everything random, and giving up on the analysis. Both are laziness of thought. My principle is clear: when the data is insufficient, the most honest answer is 'insufficient data,' not a prediction dressed up as certainty. An empty analysis is not a bad analysis — it is far more honest than one full of conclusions but empty of evidence. The problem is simply that people rarely tell the two apart, and the evidence-free version always sounds more appealing than the honest one. Every coach talks about feel. I have no feel; I have standard deviation — and standard deviation doesn't always give me an answer, but it tells me when I should stay silent. So over the coming weeks, what will I track? Not the standings, but the gap between each team's record and its true margin. Teams winning more than their net rating deserves will soon be pulled back to earth; teams losing while keeping a positive net rating will soon bounce back. This is not a random guess — it is an expectation already calibrated by probability. And if you want a certain answer right now, I'm afraid I have to disappoint you. Data is a monastery: the less noise, the more clearly you hear something trying to speak.