Betfair horse racing

The code is designed to do horse race betting on Betfair based on a trained neural network. Betting works as follows: There are two sides to every bet: back or lay. Each horse race has multiple horses, usually between 2 to 10 horses.

Overview

• Backing a horse means that you are betting that the horse will win.
• Laying a horse means that you are betting that a horse doesn't win.

The payoffs are as follows:

correctly backing: (stake * price - stake) * (1-fees) incorrect backing: -stake

Assuming a stake of $1, the payaoff is as follows:

• back correctly: price - 1
• back incorrectly: -1
• lay correctly: +1
• lay incorrectly: -price + 1

If everybody is absolutely certain that a horse will win, the betting price of that horse will be $1. Backing it will result in no payoff, nor will laying. If everybody is quite certain that horse will not win the race, the price is usually around $1000 on betfair.

The probabiliy that a horse will win can be infered from the price, and vice versa: - winning probability = 1 / price

Looking at the code the exact payoff is calculated as follows, based on the LTP (last traded price), the actual stake, and the fees. The below snipped is in historic_data_processing.py

payoff_back = np.where(self.df['winner'].values,
                       (stake * self.df['LTP'].values - stake) * (1 - fees),  # back winner
                       -np.ones((self.df['LTP'].values.shape)) * stake)  # back loser

payoff_lay = np.where(self.df['winner'].values,
                      -(stake * self.df['LTP'].values - stake),  # lay winner
                      stake * np.ones((self.df['LTP'].values.shape)) * (1 - fees))  # lay loser

The Data

The data is collected by downloading prices of each horse once per minute, starting 60 minutes before the race starts and continues to be collected until the race ends. During the race prices are collected around once every 10 seconds. Prices are last traded prices.

After the race a separate process collects the winner and writes the results to the same database.

In a separate process, the data is then transformed, so it can be fed into a neural network. The transformed data will contain the following columns:

'LTP t-0', 'LTP t-7', 'average', 'average_2d', 'back', 'countrycode', 'kurtosis', 'kurtosis_2d', 'lay', 'lay_risk', 'marketid', 'marketstarttime', 'max_2d', 'maximum', 'median', 'median_2d', 'min_2d', 'minimum', 'overrun', 'participants', 'selection_id', 'skew', 'skew_2d', 'starting_price', 'std', 'std_2d', 'winner'

Looking like that:

LTP t-0	LTP t-7	average	average_2d	back	countrycode	kurtosis	kurtosis_2d	lay	lay_risk	marketid	marketstarttime	max_2d	maximum	median	median_2d	min_2d	minimum	overrun	participants	selection_id	skew	skew_2d	starting_price	std	std_2d	winner
50.0	60.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-49.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	3893687	3.279079367849355	4.385357131797931	68.255507883	237.41174229340018	116.51357884846351	False
840.0	670.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-839.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	7262342	3.279079367849355	4.385357131797931	590.474923745	237.41174229340018	116.51357884846351	False
14.5	16.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-13.5	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	7380168	3.279079367849355	4.385357131797931	18.664683298746816	237.41174229340018	116.51357884846351	False
11.0	12.5	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-10.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	8421851	3.279079367849355	4.385357131797931	9.339514387501655	237.41174229340018	116.51357884846351	False
29.0	19.5	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-28.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	8492784	3.279079367849355	4.385357131797931	32.0	237.41174229340018	116.51357884846351	False
5.2	6.0	98.54166666666667	53.78402777777784	3.9899999999999998	GB	10.982342637974964	22.51974284698778	-4.2	-4.2	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	8869367	3.279079367849355	4.385357131797931	4.4	237.41174229340018	116.51357884846351	True
4.4	4.8	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-3.4000000000000004	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	9229409	3.279079367849355	4.385357131797931	4.6	237.41174229340018	116.51357884846351	False
32.0	32.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-31.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	10839655	3.279079367849355	4.385357131797931	40.0	237.41174229340018	116.51357884846351	False
22.0	19.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-21.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	11321256	3.279079367849355	4.385357131797931	20.893840768481603	237.41174229340018	116.51357884846351	False
8.4	6.4	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-7.4	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	11688035	3.279079367849355	4.385357131797931	7.974782176370879	237.41174229340018	116.51357884846351	False
6.0	6.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-5.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	12232392	3.279079367849355	4.385357131797931	6.6	237.41174229340018	116.51357884846351	False
160.0	250.0	98.54166666666667	53.78402777777784	-1.0	GB	10.982342637974964	22.51974284698778	0.95	-159.0	1.139233105	2018-01-24 14:40:00	990.0	840.0	18.25	16.0	4.4	4.4	1.0037970937108869	12	14838121	3.279079367849355	4.385357131797931	202.39535985152256	237.41174229340018	116.51357884846351	False
9.4	12.0	8.35	7.634861111111109	-1.0	GB	3.8745365787714015	1.3322044077460662	0.95	-8.4	1.139268720	2018-01-25 21:00:00	23.0	21.0	5.8	5.3	3.55	4.0	1.0081529125718112	6	8575987	1.9405721670660565	1.4417920045836903	7.0	6.54209446584196	4.486640644218091	False
4.0	4.1	8.35	7.634861111111109	-1.0	GB	3.8745365787714015	1.3322044077460662	0.95	-3.0	1.139268720	2018-01-25 21:00:00	23.0	21.0	5.8	5.3	3.55	4.0	1.0081529125718112	6	8706065	1.9405721670660565	1.4417920045836903	4.3	6.54209446584196	4.486640644218091	False
7.0	5.4	8.35	7.634861111111109	-1.0	GB	3.8745365787714015	1.3322044077460662	0.95	-6.0	1.139268720	2018-01-25 21:00:00	23.0	21.0	5.8	5.3	3.55	4.0	1.0081529125718112	6	10509488	1.9405721670660565	1.4417920045836903	7.4	6.54209446584196	4.486640644218091	False
4.1	4.3	8.35	7.634861111111109	2.9449999999999994	GB	3.8745365787714015	1.3322044077460662	-3.0999999999999996	-3.0999999999999996	1.139268720	2018-01-25 21:00:00	23.0	21.0	5.8	5.3	3.55	4.0	1.0081529125718112	6	11024653	1.9405721670660565	1.4417920045836903	4.280494073	6.54209446584196	4.486640644218091	True
4.6	4.7	8.35	7.634861111111109	-1.0	GB	3.8745365787714015	1.3322044077460662	0.95	-3.5999999999999996	1.139268720	2018-01-25 21:00:00	23.0	21.0	5.8	5.3	3.55	4.0	1.0081529125718112	6	11180317	1.9405721670660565	1.4417920045836903	4.562548612717727	6.54209446584196	4.486640644218091	False
21.0	21.0	8.35	7.634861111111109	-1.0	GB	3.8745365787714015	1.3322044077460662	0.95	-20.0	1.139268720	2018-01-25 21:00:00	23.0	21.0	5.8	5.3	3.55	4.0	1.0081529125718112	6	12653024	1.9405721670660565	1.4417920045836903	22.0	6.54209446584196	4.486640644218091	False
26.0	34.0	24.12	24.321888888888875	-1.0	IE	11.299591151509034	3.423915460281055	0.95	-25.0	1.139296062	2018-01-25 15:00:00	110.0	110.0	17.5	15.5	5.3	5.3	1.004864866781696	15	781222	3.2000719722913247	1.8176997877961063	25.292604111	25.129271491913286	19.541971566679	False
16.5	8.6	24.12	24.321888888888875	-1.0	IE	11.299591151509034	3.423915460281055	0.95	-15.5	1.139296062	2018-01-25 15:00:00	110.0	110.0	17.5	15.5	5.3	5.3	1.004864866781696	15	5660962	3.2000719722913247	1.8176997877961063	16.687364853692706	25.129271491913286	19.541971566679	False

The neural network input and output can then be defined as follows, where strategy is either 'lay' or 'back', depicting the payoff of the respective strategy. In other words, we are trying to predict the payoff of taking a bet on a horse, given the values defined under self.X, starting with the LTP t-0, which is the price at which we entered the bet, metrics of the distribution on how the price developed the hour before (min max, media, std skew etc).

self.X = df[['LTP t-0', 'average', 'minimum', 'maximum', 'median', 'std', 'participants', 'skew', 'kurtosis','overrun']].values
self.Y = df[['act', strategy]].values

The loss function

It's important to note that we are not trying to predict which horse wins or which one doesn't win, rather we are trying to see inefficiencies in pricing of the respective probabilities. We already know that the horse with the highest odds (lowest price) is the one most likely to win. But simply backing that horse will not yield any profit (or wil it?). This is what we're trying to find out by directly creating a loss function that calculates the payoff based on the price we enter the bet.

Starting the training

The following run will start training the neural network on a betting strategy:

app.py ts train STRATEGY CLASS COUNTRYCODES [--batchsize=<>] [--from_year=<>] [--to_year=<>] [--localhost]
app.py ts train back FlyingSpider GB,IE --from_year 2015 --to_year 2019 --localhost --batchsize=100

Structure of the code

The code consists of the follwoing packages * Collecting data from betfair and matchbook (below cronjobs need to run)

• Analyzing the pricing data of horse racing through neural networks and creating a model to make betting recommandations for backing or laying
• Executing bets according to the trained model
• Flask web interface that shows logging activity, graphical pnl overview and statistical analysis of past bets

Building the venv

You can create a venv with the environment.yml file as follows:

• Download anaconda 64 python 3
• conda create env -f environment.yml -n horse_racing (or simply run update_venv.bat). On Linux you need to use conda env create -f environment.yml
• The env will be in anaconda/envs/horse_racing

Usage:

The application is controlled over app.py. The pnl overview web server is launched over webserver.py (but better use webserver.wsgi).

Usage:
  app.py ts train STRATEGY CLASS COUNTRYCODES [--batchsize=<>] [--from_year=<>] [--to_year=<>] [--localhost]
  app.py ts backtest STRATEGY CLASS COUNTRYCODES [MODEL_PATH] [--from_year=<>] [--to_year=<>] [--localhost]
  app.py collect_prices
  app.py bet [--armed] [--sandbox_key] [--config=<>]
  app.py update_unfilled_orders [--armed] [--config=<>]
  app.py collect_results
  app.py evaluate_pnl [--overwrite_calculated_pnls] [--config=<>]
  app.py email_summary
  app.py upload_tarball FILE [DESTINATION]
  app.py map_reduce SOURCE DESTINATION CLASS [--localhost] [--use_archive]
  app.py propagate_race_results_to_price_scrape

Exmaple
  app.py ts backtest lay FlyingSpiderBookie GB,IE --from_year 2018 --to_year 2018`
  app.py ts train lay FlyingSpider GB,IE,US,NZ --localhost --from_year 2015 --to_year 2017
  app.py ts backtest lay FlyingSpider GB,IE,US,NZ --localhost --from_year 2016 --to_year 2016
  app.py propagate_race_results_to_price_scrape   adds a winner and losers column to each price in price_scrape
  app.py map_reduce price_scrape price_scrape_enriched DEBookies --localhost

Cronjobs that need to be set up:
    * * * * * sh app.sh collect_prices
    * * * * * sh app.sh bet
    * * * * * sh app.sh update_unfilled_orders --armed
    5 * * * * sh app.sh collect_results
    6 * * * * sh app.sh evaluate_pnl
    7 22 * * * sh app.sh email_summary
    9 4 * * * sh app.sh propagate_race_results_to_price_scrape
    5 5 * * * sh app.sh map_reduce price_scrape price_scrape_enriched_bookies DEBookies

Database

All data is collected onto a mongodb server which can be set up over config.ini.

Example plots of neural network training: