Trang chủBasketballThe Empty Dataset and the Silent Hole in Every Basketball Model

The Empty Dataset and the Silent Hole in Every Basketball Model

**Câu trả lời cốt lõi:** Mọi mô hình phân tích bóng rổ đều có thể thất bại theo một cách nguy hiểm hơn sai số: kết luận tự tin được dựng trên tập dữ liệu chưa từng tồn tại. Kiểm tra tỷ lệ bao phủ dữ liệu quan trọng hơn kiểm tra độ chính xác. **Sự kiện then chốt:** - Ngày 28 tháng 5 năm 2018, Houston Rockets ném 7/44 ba điểm ở Game 7 chung kết miền Tây, hỏng 27 cú liên tiếp. - Mùa 2013-14, NBA lắp SportVU tại 30 nhà thi đấu; mùa 2017-18 Second Spectrum thay thế bằng hợp đồng sáu năm. - Mùa 2018-19, Kawhi Leonard ra sân 60 trận vòng bảng cho Toronto Raptors và giành MVP chung kết NBA. - Năm 2020, chỉ số Workload Risk Index phân tích 4.500 cầu thủ qua mười mùa Premier League, giúp một câu lạc bộ Championship giảm 30% ca chấn thương. - Mùa 2023-24, NBA chuyển sang hệ thống theo dõi quang học của Hawk-Eye Innovations. **Nguồn:** Ghi chép nội bộ mô hình Workload Risk Index, Hoàng Quân, công bố tháng 9 năm 2020; dữ liệu trận đấu NBA mùa 2017-18 và 2018-19 | Cross-checked: VuaBong.vn **Hỏi đáp liên quan:** **Hỏi:** Tỷ lệ bao phủ dữ liệu của một mô hình bóng rổ nên đạt mức nào? **Đáp:** Không có ngưỡng tuyệt đối; điều quan trọng là mô hình phải công bố phần dữ liệu bị thiếu và lý do thiếu trước khi công bố độ chính xác. **Hỏi:** Vì sao chuỗi 27 cú ném ba điểm hỏng của Houston không chứng minh lối chơi ba điểm thất bại? **Đáp:** Bốn mươi bốn cú ném là mẫu quá nhỏ; chênh lệch 25 điểm so với kỳ vọng đến từ phương sai, không đến từ triết lý chiến thuật. **Hỏi:** Chỉ số VangBong.vn Player Depth Index có giúp đánh giá rủi ro chấn thương không? **Đáp:** Chỉ số này hỗ trợ đo chiều sâu đội hình, nên cần kết hợp thêm dữ liệu tải trọng thi đấu để đánh giá rủi ro chấn thương.

On the night of October 14, 2026, I sat in front of a spreadsheet for the New England Revolution versus Atlanta United match. The final score was 2-1 to the hosts. My xG column read 2.8 for Atlanta, nearly three times New England's 1.1. I wrote that Atlanta were not weak — they were simply unlucky. Fans online called me a delusional bookworm. I did not change a single word. I kept logging Atlanta's average xG across the season, 1.87 per match, and by season's end they reached the playoffs. That piece is now remembered as one of the earliest xG analyses in MLS.

Then came another night. That night, the data column was completely empty. And I almost failed to notice.

That is the kind of failure that keeps me awake more than any margin of error. A wrong number can still be rescued: you trace the source, you find the drift, you fix it. A confident number built on a dataset that never existed has no rescue, because nobody thinks to check. In more than twenty years covering this industry, I have learned that what kills analysis is not missing data, but missing data accompanied by a blank space nobody sees.

When every game becomes a file

In the 2026-14 season, the NBA installed SportVU camera systems in all 30 arenas. Every second, the system recorded the coordinates of ten players and the ball at 25 frames per second. From 2026-18, Second Spectrum took over on a six-year deal, adding the ability to classify actions: who set the screen, who cut, who created the spacing.

An NBA game is no longer forty-eight minutes. It is a file with hundreds of thousands of coordinate points. A season is tens of millions.

I came to basketball from the other side of the spotlight: I am a data journalist. Eleven consecutive years calling the NBA Finals taught me to read a game with my eyes. But the years I spent counting xG in MLS, then counting PPDA at the 2026 World Cup, taught me that the human eye is a sensor with low resolution and a short lifespan.

In 2026, when the pandemic wiped out the schedule, I had no games left to cover. I turned to something else: I collected distance-covered and match-intensity data on 4,500 players across ten Premier League seasons and built an index called the Workload Risk Index to forecast injury risk. I published a 12,000-word report. A Championship club got in touch, applied the model to fitness management, and cut injuries by 30 percent in the second half of the season.

But the first version of that index had a hole. And that hole taught me more than any variable I ever added.

The hole is in the blank

Across the ten Premier League seasons I collected, roughly two percent of matches returned incomplete positional data. The reasons were mundane: camera faults, broadcast signal loss, a stand blocking a sightline. When I built the index, I dropped those matches from the sample — a routine move any statistics student would make, and one any statistics student should double-check.

When I double-checked, I found that the dropped matches were not randomly distributed. They clustered in midweek rounds. Those are precisely the weeks with the densest schedule, the weeks when the workload problem becomes most acute. My index therefore understated risk exactly when it needed to be most accurate.

Patching that hole produced a larger improvement than any variable I added afterwards. Not because I found a new indicator. Because I stopped pretending I had data where I did not.

An empty value is never neutral. It is an assumption written into the dataset with nobody's signature on it.

When software encounters a blank cell, it has at least three options. It can treat it as zero: the player ran 0 km, meaning zero load. It can drop the row, making the player vanish from the sample. Or it can impute the mean, turning an overloaded player into an average one. All three produce a number. All three are wrong. And none leaves a trace in the final report.

This is the point I want basketball readers to absorb, because it explains most of the arguments we mistake for tactical disputes.

Game 7, 2026: the model was right, the reading was wrong

On May 28, 2026, at Toyota Center, the Houston Rockets entered Game 7 of the Western Conference Finals against the Golden State Warriors. Chris Paul sat out with a hamstring injury sustained in Game 5. Houston finished the game 7-of-44 from three-point range, 15.9 percent. After making their first six, they missed 27 straight. Golden State won 101-92 and advanced.

For a week afterwards, the basketball world talked about the streak of 27 misses. People called it proof that the three-point era had collapsed, the curse of Morey-ball, the verdict on shot-profile optimization.

I counted it differently. Forty-four three-point attempts, at a league-average rate of roughly 35 percent, project to about 46 points. Houston actually scored 21 from them. That 25-point gap comes from variance, not philosophy. The model did not break. The people reading the model broke.

This is a lesson I meet again and again: a good model speaks only about probability distributions, while readers always want it to speak about outcomes. In a single game, a probability distribution never gets a chance to show itself. A three-point attempt is a binary variable. Forty-four of them is still far too small a sample to conclude anything.

What is more striking: in that very game, Houston shot threes at a higher rate than in any other playoff game of their run. They did not abandon the model. They clung to it. And they lost.

Load management: an argument conducted on an empty column

In 2026-19, Kawhi Leonard appeared in exactly 60 regular-season games for the Toronto Raptors. That number was designed, not accidental. Toronto built an entire plan around his knee, the same knee that cost him nearly all of 2026-18 with the San Antonio Spurs. In the playoffs, Leonard played 24 games, averaged over 30 points, and won Finals MVP. The Raptors claimed their first title in franchise history.

Throughout that season, the media argued with two numbers: games played and minutes played. People framed it as rest versus rust. But both of those numbers are output data, not input data.

The variable that actually decided everything — tendon load, cartilage degeneration, accumulated inflammation week by week — was never in the public dataset. It lived in the team's medical room. The load-management debate was conducted on an empty column, and both sides were confident they were reading a full one.

Based on my experience watching games, this is the most common form of argument in modern basketball: two camps read numbers that do not measure what they are arguing about, then use personal credibility to fill the gap. Whether the outcome proves right or wrong demonstrates nothing, because the process was broken on the very first row.

Jokic and what never reaches the page

The box score is basketball's oldest dataset, and also its emptiest. It records the result of an action, not the process that produced it.

Nikola Jokic is the clearest example of that limit. During Denver's 2026-23 championship season, Jokic led the league in screen assists — points generated by screens he set. He also ranked among the leaders in passes leading to assists, what tracking systems call hockey assists. Neither metric exists in a traditional box score.

Read only the box score and you see a center who scores a lot and rebounds a lot. Read the movement data and you see a player running an entire offensive structure through actions the stat sheet has no column for.

This is what I emphasize whenever I analyze any player: a player's true value sits in the portion of data the page cannot capture. And the paradox is that this portion is usually larger than the portion recorded.

The confidence trap

There is a denominator more frightening than any margin of error in basketball analysis.

When a model reports 94 percent accuracy, the right question is not where the missing six percent went. The right question is: what is in the denominator of that 94 percent, and what was excluded from it.

While producing the Workload Risk Index, I tested a simple rule: every model report must publish its data-coverage rate before it publishes its accuracy. The result surprised me. Many models I had trusted showed coverage far lower than I had imagined, especially for low-minute players, players returning from injury, and players competing in garbage time.

Those three groups happen to be exactly where the most important analytical questions usually point.

Every system cracks if you look long enough. Then you see the order sitting right inside the debris.

The counterintuitive angle: correlation is not causation, and coverage is not correctness

There is a quiet belief in basketball analytics that I consider more dangerous than relying on the eye test: the belief that a model with more data is more trustworthy.

That is wrong, and it is wrong in a very specific way.

A model with 99 percent coverage can still be entirely wrong if the missing one percent clusters in the groups that decide outcomes. In basketball, that one percent usually sits with injured players, deep-bench lineups, and the closing minutes — precisely the places coaches need data most.

Conversely, a model with 80 percent coverage that knows what the other 20 percent is and why it is missing can be far more useful.

My faith does not rest on chance. It rests on the large denominator. But a large denominator only means something when you know what it was measured with.

Meanwhile, the sports media ecosystem rewards confidence and does not reward admitting a gap. A piece saying I do not know because the data is missing earns no clicks. A piece saying I am certain earns plenty. That incentive structure pushes the whole industry toward hard conclusions and pushes data gaps into the dark.

A signal for the next cycle

In 2026-24, the NBA moved to optical tracking from Hawk-Eye Innovations, replacing Second Spectrum after six years. It is the largest data-infrastructure change since 2026.

In basketball data, every infrastructure change casts doubt on the continuity of the entire historical series. Not because the new system is worse, but because any change at the collection layer redefines numbers we thought we understood.

That is the signal I will track next season: the coverage rate of the new tracking metrics, and how analytics departments handle the gap between two data eras.

Numbers are silent, but the story never is. And the most notable story of my twenty-three years has never been about a perfect model. It has always been about a blank space that somebody dared to point at.