Recently I’ve found an interesting paper that predicts winning NBA teams for each game. Despite looking for betting simulations or experiments where it were shown how these predictions relate to some baseline value and how they can (not) generate consistent profits, such information was nowhere to be found.
What is more, this is certainly not the first scientific paper with a lack of betting simulations to evaluate profitability of a certain method. And I’ve seen plenty – they report model accuracy, e. g. the winner is correctly predicted in 60 % of times for season 2017/2018 – but you as a reader don’t really know if this can help you earn money from betting.
The baseline model that you’re trying to beat is bookmaker probabilities. You take bookie’s odds and transform them to probabilities. If you’re able to consistently outperform the bookie, you’re all set to earn profit consistently. (By the way, proper money management should also be used.) However, that is much easier said than done.
Nothing else remains but to see how bookies performed in the previous NBA seasons in terms of prediction accuracy, so in the future both you and I know if some method researchers try to sell is any good and worth looking into.
| season | number of games in regular season | bookie prediction accuracy |
| 2016/2017 | 1230 | 66.3 % |
| 2017/2018 | 1230 | 68.5 % |
| 2018/2019 | 1230 | 67.3 % |
| 2019/2020 | 1059 | 66.8 % |
| 2020/2021 | 1080 | 65.9 % |
| 2021/2022 | 1230 | 67.8 % |
| 2022/2023 | 776* | 64.0 % |
Compare the table above with some prediction accuracies from the literature:
- 63 % for 2016/2017, 67 % for 2017/2018, 66 % for season 2018/2019,
- under 60 %, even though not clear which season, probably 2016/2017,
- range 65 – 68 % for first half of season season 2017-2018, but paper does not say whether random shuffle or time series split was used to perform a train/test split.
Are you surprised when comparing researchers’ predictions and bookie outcomes? After a couple of years of hobby projects trying to beat the bookie, I’m not. Bookies predictions for big leagues are very good and difficult to beat in the long run, especially using automated approaches only.
Truth be told, while researching for numbers above, some interesting papers popped up on my radar with reported prediction accuracies of 70 % and above. I can’t lie, this looks very interesting, but stuff for another post.
Before ending this post, a question: anyone playing daily fantasy sports? Providers of player fantasy points predictions such as DailyFantasyFuel also lack metrics to track past performance, so you don’t really know what you’re using. Sure, use those projections when building your lineup and you soon find out if they’ll help you earn an extra buck or not, but wouldn’t it be nice to know what to expect beforehand?
Anyone interested in a similar analysis for your preferred league or online source of projections?