How To Predict Winners With BitOdds Data and Python

We can make some observations about the dataset from this sample of rows:

1. There is a row for each bet that can be made. For example, there is one row for betting on Suns to win in Suns vs Hawks and one row for betting on Hawks to win.
2. The market type is indicated in the market column and takes on the values event_winner , spread and total .
3. For spread and total markets, the market_qualifier column indicates the spread value or total value associated with the selection. For example, in the row with index 2 the market_qualifier indicates the row is for Suns -5.5 points.
4. We have the market odds for 6 sportsbooks in the columns labelled sportsbet_odds , cloudbet_odds etc and the average of the sportsbooks’ odds in the average column.
5. We have an indication of whether the selection was a winning bet or not in the is_winner column and the final score in the result column.

We use selections.shape to reveal the DataFrame has a total of 7,070 rows and 17 columns and we generate some statistics on the numerical columns using selections.describe() :

From this output we observe:

1. There are roughly 7,000 odds for most sportsbooks but only 210 for Trust Dice. The missing odds appear in the DataFrame as NaN .
2. The odds range from 1.03 to 17.33 and the average is around 2.1.

We can also visualize this data by plotting histograms of each column with selections.hist() :

Again we see that the odds range between 1 and 17 but we now see that almost all odds are below 5.

To get an idea of the values taken by the columns containing text, we can use the value_counts() method on the column. For example, to see the range of values in the market column we call selections.market.value_counts() :

This indicates that there are 2,370 rows with market set to total , 2,358 with spread and 2,342 with event_winner .

Biggest upsets of the season

Suppose we want the biggest underdog wins. We first massage the data a little bit:

The code above:

1. Drops all the rows except where market is equal to 'event_winner' .
2. Changes the values in the is_winner column from text strings of "yes" and "no" to boolean values of True and False .
3. Adds a new column side which will be "home" or "away" .
4. Adds a new column best_odds which will be the highest odds offered by any sportsbook.
5. Renames the columns average and date_utc columns to average_odds and date .
6. Drops all the columns except for home , away , date , result , selection , side , average_odds , best_odds and is_winner .

The DataFrame now looks like:

Finding the biggest upsets can now be achieved by:

1. Filtering the DataFrame to only contain rows where is_winner is True .
2. Sorting the resulting DataFrame by odds in descending order.
3. Keeping only the top 5 selections.

Pandas make it easy to do all this in one line of code:

So the biggest upset of the season was the Houston Rockets beating the Milwaukee Bucks where the Rockets had average odds of 8.46. Interestingly the top three upsets were all against the Bucks!

Is the home team advantage accurately reflected in the odds?

To find out if there was a home team advantage in the NBA 2020-21 season we can compare the rate at which home teams win with the rate at which away teams win. To calculate the win rates we can use a neat trick: when you tell Pandas to calculate the average of a column of booleans using the mean() function it will treat False as 0 and True as 1 . So if we take the mean of the is_winner column and it contained 8 True values and 2 False values we would get 0.8 or 80% which is exactly the win rate.

Now we want to calculate the win rate separately for the rows where side = "home" and where side = "away" . We can achieve this by telling Pandas to groupby("side") before calculating the mean on is_winner :

This reveals that away teams won only 45.3% of games whereas home teams won 54.7% of games. This is a whopping 9.4% difference! Let’s see whether this difference is priced into the odds…

The average odds for home teams is indeed shorter than for away teams but this doesn’t reveal whether the sportsbooks have adjusted the odds by the correct proportion given the magnitude of the home side advantage. To determine this we convert the odds into implied probability which is the probability a bet would have if the odds were fair (ie if the sportsbook edge was zero). We calculate implied probability as 1/odds :

Now we calculate the average implied probability for home versus away bets:

We see that the difference in average implied probability values are in line with the win rates we calculated. The implied probabilities are a few percentage points higher because of the sportsbook edge. Unfortunately, this means we can’t use our estimate of the home-side advantage alone to beat the sportsbooks.

Do teams perform better after a win?

Another interesting question to ask is whether teams perform better coming off a win vs coming off a loss. To investigate this we first need to add an extra column to our DataFrame indicating whether the team had won their previous game. We can achieve this with the following code:

Here process the selections for each team individually, using the shift() method to translate the is_winner values down one row, storing them in a new column called won_last_game . Once we have performed this update on separate DataFrames for each team we recombine the DataFrames using pd.concat() . The result is:

We see that the won_last_game value always matches the previous row’s is_winner value except for the team’s first game of the season where won_last_game is NaN . We can’t use these rows where won_last_game is NaN so we drop them with:

A naive look at the impact of coming off a win or a loss would be to calculate the win rate based on whether or not the team won their last game:

This indeed shows that teams coming off a win do win their next game more often (54%) than teams coming off a loss (46%). However, this might not be causal as there is a confounding factor we need to consider: good teams will more often be coming off a win and will win their next game because they are good, whereas bad teams will more often be coming off a loss and lose their next game because they are bad.

To compensate for this we can examine the performance of favourites vs underdogs based on their previous results. We first add a new column indicating whether a team is an underdog based on whether their odds are over 2:

We again calculate the win rate but this time based on won_last_game and is_underdog :

In this table, the top-left cell is the win rate for favourites that didn’t win their last game ( is_underdog = False , won_last_game = False ), the top-right cell is the win rate for underdogs that didn’t win their last game ( is_underdog = True , won_last_game = False ) etc.

We see that when a team is the favourite, having won their last game only increases their chance of winning by 2% (from 64% to 66%). However, for underdogs, the effect is much larger. An underdog coming off a win is 5% more likely to win than an underdog coming off a loss (from 30% to 35%).

Has this difference been priced into the odds?

This table shows the average implied probability for the different combinations of is_underdog and won_last_game . If the impact of winning the last game was priced into the odds, these values would be the same as the win rate plus a few percentage points due to the sportsbook edge. We see that the sportsbooks’ implied probability is around 2% higher than the win rate in all cases except for underdogs coming off a win, where it is only 0.7%. This could indicate the sportsbooks underestimated the increase in underdogs’ chances when coming off a win.

Predicting winners and finding profitable bets

Now that we have identified some variables that can be used to predict which team will win, we are ready to build our prediction model. There are many excellent machine learning and statistics frameworks for Python, including:

• scikit-learn
• statsmodels
• TensorFlow
• PyTorch

We will use scikit-learn as it is both easy to use and very powerful. The first step is to decide on a model class. Scikit-learn has a large number of these such as ordinary least squares, nearest neighbours, random forests and neural networks. Finding the best model type to use is a bit of trial and error, although scikit-learn does have tools to help with model selection. We found that the support vector machine model is well suited to our task so we start by importing the SVM class:

Scikit-learn can only handle numeric data so we cannot use our side the column which contains the text values "home" and "away" . Instead, we create a boolean column named is_home . Next, we need to split our data into a training subset and an evaluation subset. We pretend that we decided to train our model exactly halfway through the season and then use the model for betting for the remaining half of the season. We, therefore, need to sort the selections by date and then split the DataFrame into two halves using the np.split() function. Finally, we assign the values we want to use as inputs to the model to the variables X_train and X_test and the is_winner values we want to predict to y_train and y_test .

Now we are ready to train our model. This is as simple as calling the fit() method on the model. Once the model is trained we can call predict() on it with the validation data to check it’s accuracy.

So our model correctly predicts the winner 69.9% of the time. This seems pretty good, but is it enough to make a profit betting with these predictions? To find out we should first determine which selections we would bet on. We want to place a bet whenever the predicted probability multiplied by the odds is greater than 1. We can predict the probability of each selection winning by using the predict_proba() method on our model. Using this we find the selections recommended by our model and then calculate what the profit would have been if we had bet on these:

Finally, we can calculate our return on investment by adding up the profit column and dividing by the number of bets:

So our model would have generated a very decent 5% ROI.

Since our model only depends on three boolean inputs, we can explore what it predicts the probability of a win is and what odds we should bet on for all combinations of these inputs:

We see from this output that the model only really cares about the is_underdog value. It predicts underdogs have a 35.91% probability of winning regardless of whether they are the home team or whether they won their last game. Similarly, the model predicts favourites have a 62.76% chance of winning in all cases. This means we should bet on any teams that have odds between 1.593 and 1.999 or odds over 2.785. It is hard to believe this model would work in general but at least in the last half of the 2020-21 season, it would have returned a 5% profit.

Next steps

In this article, we have shown how to build a simple prediction model with BitOdds data and sci-kit-learn. Based on the 2020-21 season data it appeared to generate a very respectable 5% profit. The next step would be to test the robustness of this model. This could be done by using sci-kit-learn’s cross-validation functionality or by exporting the data from a different NBA season from BitOdds and applying the model to that.

To develop the model even further you could incorporate additional information such as the teams playing, the win rates of the teams etc. You could also try building a model that predicts the score of the game and use this to bet on the total points and spread markets.

With large amounts of sports data freely available and incredible open-source tools like Python, Jupyter, Pandas and sci-kit-learn, there has never been a better time to try your hand at data-driven sports betting. Good luck!