Linear regression from scratch: Predicting F1 results

Before we start¶

This notebook can be browsed interactively on Kaggle: https://www.kaggle.com/dorin131/f1-predictions-blog

Prologue¶

The other day I learned how to implement linear regression and have been itching to make good use of this newly acquired knowledge. It's funny how ML can be applied to almost anything but when it came to actually picking one thing, I didn't know what. What would be fun? What is something that I'm interested in? So I start going through the Kaggle dataset catalogue... and, BINGO! F1 results from 1950 to 2020! What can I do with this? What else than try to predict who's going to win the next race? I suspect it may not the best fit for a linear regression model but it will be fun.

The plan¶

So, the plan is to combine the tables provided in this dataset and end up with a few features that I think may have the biggest effect on the finish position of a driver. Let's say...

• driver standing
• constructor standing
• driver grid position

If only it was that simple to predict the position, right? Well, it's not that simple and our predictions won't be all that accurate, but who cares!

Below you'll see me load the data, do some transformations, joins, filtering, conversions, etc. Once all of this boring part is done, we'll get into implementing the model and then training it!

In [7]:

# Importing some libraries we're going to need
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import os

from pandas.plotting import scatter_matrix

Getting the data¶

In [25]:

# Looking at what files we've got and getting their paths
for dirname, _, filenames in os.walk('/kaggle/input'):
    for filename in filenames:
        print(os.path.join(dirname, filename))

/kaggle/input/formula-1-world-championship-1950-2020/races.csv
/kaggle/input/formula-1-world-championship-1950-2020/constructor_results.csv
/kaggle/input/formula-1-world-championship-1950-2020/drivers.csv
/kaggle/input/formula-1-world-championship-1950-2020/constructors.csv
/kaggle/input/formula-1-world-championship-1950-2020/lap_times.csv
/kaggle/input/formula-1-world-championship-1950-2020/status.csv
/kaggle/input/formula-1-world-championship-1950-2020/driver_standings.csv
/kaggle/input/formula-1-world-championship-1950-2020/seasons.csv
/kaggle/input/formula-1-world-championship-1950-2020/pit_stops.csv
/kaggle/input/formula-1-world-championship-1950-2020/sprint_results.csv
/kaggle/input/formula-1-world-championship-1950-2020/constructor_standings.csv
/kaggle/input/formula-1-world-championship-1950-2020/results.csv
/kaggle/input/formula-1-world-championship-1950-2020/circuits.csv
/kaggle/input/formula-1-world-championship-1950-2020/qualifying.csv

In [26]:

results = pd.read_csv('/kaggle/input/formula-1-world-championship-1950-2020/results.csv')
driver_standings = pd.read_csv('/kaggle/input/formula-1-world-championship-1950-2020/driver_standings.csv')
constructor_standings = pd.read_csv('/kaggle/input/formula-1-world-championship-1950-2020/constructor_standings.csv')

Quick look at the data¶

We're just going to print the first 5 rows of each dataset and check the types

In [27]:

results.head()

Out[27]:

	resultId	raceId	driverId	constructorId	number	grid	position	positionText	positionOrder	points	laps	time	milliseconds	fastestLap	rank	fastestLapTime	fastestLapSpeed	statusId
0	1	18	1	1	22	1	1	1	1	10.0	58	1:34:50.616	5690616	39	2	1:27.452	218.300	1
1	2	18	2	2	3	5	2	2	2	8.0	58	+5.478	5696094	41	3	1:27.739	217.586	1
2	3	18	3	3	7	7	3	3	3	6.0	58	+8.163	5698779	41	5	1:28.090	216.719	1
3	4	18	4	4	5	11	4	4	4	5.0	58	+17.181	5707797	58	7	1:28.603	215.464	1
4	5	18	5	1	23	3	5	5	5	4.0	58	+18.014	5708630	43	1	1:27.418	218.385	1

In [28]:

results.dtypes

Out[28]:

resultId             int64
raceId               int64
driverId             int64
constructorId        int64
number              object
grid                 int64
position            object
positionText        object
positionOrder        int64
points             float64
laps                 int64
time                object
milliseconds        object
fastestLap          object
rank                object
fastestLapTime      object
fastestLapSpeed     object
statusId             int64
dtype: object

In [29]:

driver_standings.head()

Out[29]:

	driverStandingsId	raceId	driverId	points	position	positionText	wins
0	1	18	1	10.0	1	1	1
1	2	18	2	8.0	2	2	0
2	3	18	3	6.0	3	3	0
3	4	18	4	5.0	4	4	0
4	5	18	5	4.0	5	5	0

In [30]:

driver_standings.dtypes

Out[30]:

driverStandingsId      int64
raceId                 int64
driverId               int64
points               float64
position               int64
positionText          object
wins                   int64
dtype: object

In [31]:

constructor_standings.head()

Out[31]:

	constructorStandingsId	raceId	constructorId	points	position	positionText	wins
0	1	18	1	14.0	1	1	1
1	2	18	2	8.0	3	3	0
2	3	18	3	9.0	2	2	0
3	4	18	4	5.0	4	4	0
4	5	18	5	2.0	5	5	0

In [32]:

constructor_standings.dtypes

Out[32]:

constructorStandingsId      int64
raceId                      int64
constructorId               int64
points                    float64
position                    int64
positionText               object
wins                        int64
dtype: object

Dropping the columns we don't need¶

In [33]:

# We only need a few columns, which we're getting here
results = results[["raceId", "driverId", "constructorId", "grid", "position"]]
results.head()

Out[33]:

	raceId	driverId	constructorId	grid	position
0	18	1	1	1	1
1	18	2	2	5	2
2	18	3	3	7	3
3	18	4	4	11	4
4	18	5	1	3	5

In [34]:

driver_standings = driver_standings[["raceId", "driverId", "position"]]
# Rename the "position" column do avoid conflict with the "position" column from results.csv
driver_standings = driver_standings.rename(columns={"position": "driverStanding"})
# Use current driver standings for the next race
driver_standings["raceId"] += 1
driver_standings.head()

Out[34]:

	raceId	driverId	driverStanding
0	19	1	1
1	19	2	2
2	19	3	3
3	19	4	4
4	19	5	5

In [35]:

# Again, picking the columns we need and renaming "position"
constructor_standings = constructor_standings[["raceId", "constructorId", "position"]]
constructor_standings = constructor_standings.rename(columns={"position": "constructorStanding"})
# Use current constructor standings for the next race
constructor_standings["raceId"] += 1
constructor_standings.head()

Out[35]:

	raceId	constructorId	constructorStanding
0	19	1	1
1	19	2	3
2	19	3	2
3	19	4	4
4	19	5	5

Joining the data¶

In [36]:

# Joining results with driver standings. This will add the "driverPosition" column to our results
results_driver_standings = pd.merge(results, driver_standings, on=["raceId", "driverId"], how="inner")
results_driver_standings.head()

Out[36]:

	raceId	driverId	constructorId	grid	position	driverStanding
0	18	1	1	1	1	5
1	18	2	2	5	2	13
2	18	3	3	7	3	7
3	18	4	4	11	4	9
4	18	5	1	3	5	12

In [37]:

# Now we join the constructor standings and we end up with everything we need in one place
joined_data = pd.merge(results_driver_standings, constructor_standings, on=["raceId", "constructorId"], how="inner")
joined_data.head()

Out[37]:

	raceId	driverId	constructorId	grid	position	driverStanding	constructorStanding
0	18	1	1	1	1	5	3
1	18	5	1	3	5	12	3
2	18	2	2	5	2	13	6
3	18	9	2	2	\N	14	6
4	18	3	3	7	3	7	7

Sense checking the data¶

Things look good so far but I've got no idea whether the numbers are correct and I haven't messed up something. To verify things, I'm going to print out the results for the last couple of races in 2020 and compare with F1 results from Wikipedia.

In [38]:

joined_data.sort_values(by='raceId', ascending=False).head(60)

Out[38]:

	raceId	driverId	constructorId	grid	position	driverStanding	constructorStanding
22157	1096	825	210	16	17	13	8
22147	1096	4	214	10	\N	9	4
22138	1096	830	9	1	1	1	1
22139	1096	815	9	2	3	3	1
22140	1096	844	6	3	2	2	2
22141	1096	832	6	4	4	6	2
22142	1096	847	131	6	5	4	3
22144	1096	846	1	7	6	7	5
22145	1096	817	1	13	9	12	5
22146	1096	839	214	8	7	8	4
22143	1096	1	131	5	18	5	3
22148	1096	840	117	14	8	15	7
22153	1096	822	51	18	15	10	6
22149	1096	20	117	9	10	11	7
22155	1096	849	3	20	19	20	10
22154	1096	848	3	19	13	19	10
22156	1096	854	210	12	16	16	8
22152	1096	855	51	15	12	18	6
22151	1096	842	213	17	14	14	9
22150	1096	852	213	11	11	17	9
22127	1095	855	51	13	12	18	6
22118	1095	847	131	1	1	4	3
22119	1095	1	131	2	2	5	3
22120	1095	832	6	7	3	6	2
22121	1095	844	6	5	4	3	2
22122	1095	4	214	17	5	9	4
22123	1095	839	214	16	8	8	4
22124	1095	830	9	3	6	1	1
22125	1095	815	9	4	7	2	1
22126	1095	822	51	14	9	10	6
22133	1095	852	213	0	17	17	9
22128	1095	840	117	15	10	15	7
22134	1095	848	3	19	15	19	10
22129	1095	20	117	9	11	11	7
22136	1095	846	1	6	\N	7	5
22135	1095	849	3	18	16	20	10
22137	1095	817	1	11	\N	12	5
22132	1095	842	213	10	14	14	9
22131	1095	825	210	8	\N	13	8
22130	1095	854	210	12	13	16	8
22107	1094	4	214	9	19	9	4
22098	1094	830	9	1	1	1	1
22099	1094	815	9	4	3	3	1
22100	1094	1	131	3	2	6	3
22101	1094	847	131	2	4	4	3
22102	1094	832	6	5	5	5	2
22103	1094	844	6	7	6	2	2
22104	1094	817	1	11	7	12	5
22105	1094	846	1	8	9	7	5
22106	1094	839	214	10	8	8	4
22113	1094	849	3	18	18	20	10
22108	1094	822	51	6	10	10	6
22114	1094	20	117	16	14	11	7
22109	1094	855	51	12	13	18	6
22116	1094	854	210	15	16	16	8
22115	1094	840	117	20	15	15	7
22117	1094	825	210	19	17	13	8
22112	1094	848	3	17	12	19	10
22111	1094	852	213	13	\N	17	9
22110	1094	842	213	14	11	14	9

I picked George Russell as my point of reference. Why him do you ask? Maybe because I walked past him on the street once, so that makes him the F1 driver I came closest to. Good enough reason for me.

We can see below that in the Abu Dhabi 2022 race (1096), George Russell (847) came 5th, starting 6th on the grid. That looks right.

(https://en.wikipedia.org/wiki/2022_Abu_Dhabi_Grand_Prix)

Now to check that the driver and constructor standings are correct, we have to look at the previous race. At the end of the Sao Paolo race, he was 4th in the driver standings and Mercedes was 3rd in constructors'.

(https://en.wikipedia.org/wiki/2022_S%C3%A3o_Paulo_Grand_Prix)

Great! It all checks out. We can continue.

In [39]:

# Let's see how many examples do did we end up with
len(joined_data)

Out[39]:

22158

Quick plot¶

In [40]:

joined_data[["grid", "driverStanding", "constructorStanding", "position"]].hist(figsize=(12, 8))
plt.show()