Formula One Rating System Walk-Through, Part II

This is the second part of the deep dive into my Formula One model. Part I provided a high-level overview of the approach. Today's post will go into the details of the implementation, how the model calculates reliability and performance for both teams and drivers, and allocates results to each. Future posts will outline how the knobs are tuned, evaluate its predictive power, and walk through how the predictive ratings are aggregated into backwards-looking "resume" ratings like the ones used in the 2021 preview.

Calculating Reliability

As stated in the previous post, the goal of our reliability model is to predict the odds of failure in an average kilometer for each car and for each driver. The model tracks reliability in five buckets:

1. Per-driver reliability (ability to not crash);
2. Aggregate reliability of all drivers in the field;
3. Per-team reliability;
4. Aggregate reliability of "mature" teams in the field; and
5. Aggregate reliability of "new" teams in the field.

These are measured in terms of “success kilometers” and “failure kilometers”. The probability of per-kilometer success of a car or driver is simply [(success kilometers) / (total kilometers)]. Since the probability of success each kilometer is independent of the previous kilometer, the probability of a car not suffering a mechanical DNF in an X-km race is the per-kilometer car success probability raised to the Xth power; the same calculation is also done using the driver per-kilometer success probability. Since driver-related DNFs are independent of car-related DNFs, the probability of an entrant successfully completing the whole race distance is (car success probability * driver success probability).

We separate out the "new" team reliability from the "mature" team reliability due to the chaotic nature of the early days of Formula One. There have been just shy of 500 unique teams to take part in F1 over the decades. Of those, around 300 joined in the 1950s, another 100 joined in the 1960s, and the rest started their lineage in 1970 or later. While there is a big gap between today's "haves" and "have-nots", the gap was noticably larger in the early days of the sport. For example, the 1953 German Grand Prix alone had more one-race teams-slash-privateers-as-constructors (Dora Greifzu, Rennkollektiv EMW, Ernst Loof, Erwin Bauer, Gunther Bechem, Oswald Karch, and Rodney Nuckey) than truly new teams have joined Formula One in the last 20 years (Toyota, Super Aguri, HRT, Team Lotus, Virgin, and Haas).

For each entrant there are three possible outcomes in an X kilometers long race. If the entrant:

1. completes the entire race distance, we update the number of success kilometers for four buckets -- the first three buckets plus one for either the "new" or "mature' team -- by X;
2. completes Y kilometers but then suffers a mechanical DNF, we:
1. add Y success kilometers to those four buckets; and
2. add a small constant in the range [0, 1] to the number of failure kilometers for the two relevant team-related buckets (team_reliability_failure_constant);
3. completes Z kilometers but then suffers a driver-attributed DNF, we:
1. add Z success kilometers to the four relevant buckets; and
2. add a small constant in the range [0, 1] to the number of failure kilometers for the two driver-related buckets (driver_reliability_failure_constant).

The driver_reliability_failure_constant and team_reliability_failure_constant parameters allow us to hone in on the best ratios to predict driver and car failure.

Before adding the results of the current race to any of the buckets, we apply a decay factor to the existing data (driver_reliability_decay or team_reliability_decay) to gradually age out existing data.

Before the first race of each season we regress the aggregate "new" team reliability back to the "mature" team reliability by a certain amount. Before the first race a driver or team participates in in a new season, we regress their reliability back to the mean by a fixed percent (driver_reliability_regress and team_reliability_regress). Whether we regress the team to the "new" or "mature" bucket depends on the number of races in which the team has participated. At this point we also cap the total kilometers in their reliability data set to a fixed multiple of an average race distance (driver_reliability_lookback and team_reliability_lookback). This makes the reliability metrics slightly more sensitive to change during the earlier stages of the season.

When a new driver or team appears for the first time, we give them the default reliability rate of the entire field.

While we could use an existing implementation of a Kaplan-Meier survival estimator or an exponential survival model in something like Scikit or Lifelines, these approaches are heavyweight and did not produce a statistically superior predictor. Additionally, they increased the runtime of our model from ~90 seconds to 30 minutes (Kaplan-Meier) or 180 minutes (exponential survival).

Calculating Performance

Performance is calculated using a hybrid Elo model, in which drivers and teams are modeled independently and then combined for each entrant. Elo ratings are updated for each entrant after each qualifying session, and in races only if an entrant does not suffer a driver-related or car-related DNF.

New drivers start with driver_elo_initial points, and new teams with team_elo_initial points.

Car vs Driver

The combined entrant rating and K-Factors are calculated using a weighted average of the driver Elo information and the car (team) Elo information. This weighting can change over time, and is specified by the team_share_spec parameter. The parameter is specified as [InitialTeamShare]_[YearWidth]_[StepHeight], where InitialTeamShare is the percent of the entrant Elo information coming from the car in 1950 (the first season), and then every YearWidth seasons that number will increase by StepHeight percent. For example, a spec of 50_4_1 means that from 1950 through 1953, the team contributes 50% of the Elo information to the entrant Elo information, then from 1954 through 1957 it contributes 51%, 1958 through 1961 is 52%, and so on. A spec of [N]_0_0 signals a constant share of N% throughout the history of Formula One. This allows us to empirically test the hypothesis that the share of overall results due to the car has steadily increased over time, without risking overfitting on a per-season basis.

We apply this weighted average between car and driver to both the Elo rating and K-Factor when creating the combined Elo information. For example, in a season where the car accounts for 60% of the outcome, a combination of the following car and driver would produce these combined Elo and K-Factor ratings:

	Car	Driver	Combined
Elo	1260	1560	756 + 624 = 1380
K-Factor	13	18	7.8 + 7.2 = 15.0

This gives us the basic combined Elo rating and K-Factor of the entrant.

Starting Position Advantage

For races (but not qualifying) we must also take into account starting position on the grid. The model treats starting position much like home field advantage, in that it will give the driver closer to the front of the grid a boost in their Elo ratings for that one head-to-head prediction. It is likely that this advantage is non-linear, meaning that at some point there is no significant difference between starting N positions ahead of another entrant versus N+1 positions ahead; e.g., there may be a noticeable difference between having a 2 space advantage versus a 5 space advantage, but less difference between 12 spaces and 15 space. It is also possible that the base advantage of a single grid spot has increased over time, contributing to the sense that Formula One races are glorified parades.

We model this advantage through a combination of two parameters:

• position_base_spec: formatted the same as team_share_spec, this allows us to vary the number of Elo points a single grid spot confers as an advantage over time; and
• position_base_factor: a value used as the ratio for a geometric sum, mapping the number of grid positions to a multiplier of the base spec for that year.

Plugging all this into the geometric sum formula, the Elo advantage E_A conferred by G grid positions in a season with a base Elo grid advantage of E_B and a factor of F is:

E_A = E_B * [(1 - F^G) / (1 - F)]

The values of the base advantage and the base factor control the shape of the curve for this advantage. Values of F closer to 1 create a more linear shape, while lesser values create a flatter shape.

Predicting Performance Outcomes

Once we have the combined Elo rating and (for races) the start position advantage, we can then use these numbers to predict the probability that one entrant will finish in front of another entrant (assuming both finish) using the expected score logistic equation.

For this equation, though, we need to determine the correct denominator. From the Wiki page, a denominator of 400 means that “for each 400 rating points of advantage over the opponent, the expected score is magnified ten times in comparison to the opponent's expected score”. Or, in terms of odds, a denominator of X means that an Elo rating advantage of X points represents a 10-to-1 favorite, whereas a rating advantage of 2X points represents a 100-to-1 favorite.

Since qualifying is shorter and has less variance, we may expect that the same performance difference in qualifying would yield greater odds of winning than the same performance difference in a race. The model allows us to specify elo_exponent_denominator_qualifying and elo_exponent_denominator_race separately in order to keep the Elo rating constant between event types, but still capture the differences in these types.

Combined Head-to-Head Model

Putting this all together the probability that driver A in car X finishes (entrant E) ahead of driver B in car Y (entrant F) is:

• the probability that entrant E finishes the race but F does not (the car and driver reliability calculations); plus
• the probability that entrant E does not finish the race but completes more laps than F (per-lap reliability calculations); plus
• conditional on both E and F completing the race, the probability that E outperforms F (performance calculations).

Of those, the second is the most complex to calculate, but contributes the least to the final probability.

If both entrants complete the race (or participate in the qualifying session) we must reallocate Elo points. The K-Factor used for any transfer of points is the average of the combined K-Factor for each entrant. The Elo rating for each entrant is the combined Elo rating of each entrant, plus the starting position advantage points for whichever entrant starts first.

For example, let's say that there are four entrants which complete a race, two teams of two drivers each. In this year the car accounts for 60% of the performance, the base Elo advantage for one grid spot is 20 points, and the position factor is 0.75.

The two teams:

Team	Elo	K-Factor	Reliability
X	1400	16	93%
Y	1350	20	91%

The four drivers:

Driver	Elo	K-Factor	Reliability
A	1400	12	98%
B	1300	20	90%
C	1525	10	95%
D	1350	16	91%

The four entrants, in grid order:

Entrant	Elo	K-Factor	Reliability	Grid
E: A+X	1400	14.4	91.1%	1
G: C+Y	1420	16.0	86.4%	2
H: D+Y	1350	18.4	82.8%	3
F: B+X	1360	17.6	83.7%	4

The head-to-head performance-only probabilities, assuming an Elo denominator of 240:

Entrant 1			Entrant 2		E1
Name	Elo	Grid	Name	Elo	WinProb
E	1400	20	G	1420	50.0%
E	1400	35	H	1350	69.3%
E	1400	46	F	1360	69.5%
G	1420	20	H	1350	70.3%
G	1420	35	F	1360	71.3%
H	1350	20	F	1360	52.4%

Note that without the one spot grid advantage for E over G, G would be the slight favorite, whereas on performance now it's a dead heat.

The probabilities above are also conditional on both entrants finishing. Digging into the E vs G matchup a bit more, there are the following scenarios, along with the probability that E finishes ahead of G:

Scenario	P(happening)	P(E wins) if this happens	P(E wins) total
Both finish	78.7%	50.0%	39.4%
E finishes G DNFs	12.4%	100.0%	12.4%
E DNFs G finishes	7.7%	0.0%	0.0%
Double DNF	1.2%	50%	0.6%
Overall			52.4%

Putting it all together, a much quicker driver becomes the slight underdog against an average-yet-reliable driver in a solid-but-slightly-more-reliable car who's managed to qualify on pole.

Coming up...

Part III will discuss its predictive performance. Part IV will discuss how predictions get aggregated into metrics which span one or more year.

Formula One Rating System Walk-Through, Part I

New year, new energy to write up some content about the model at the heart of my Formula One ranking and prediction system. This is the model used in a few posts on FiveThirtyEight, including the 2021 season preview and (brief) season retrospective. It's also an evolution of the model used in the 2018 "Best Formula One Driver of All Time" article, which was the target of some (reasonable) criticism.

Over the next few posts I'll go into the conceptual parts of the model, describe how it's implemented, outline how the knobs are tuned, evaluate its predictive power, and walk through how the predictive ratings are aggregated into backwards-looking "resume" ratings like the ones used in the 2021 preview.

Overview

At its heart the objective of this system is to predict the answer to “will driver A in car X be in front of driver B in car Y at the end of the event?” It attempts to optimize this predictive model for both qualifying sessions and races, equally taking into account all eras of Formula One (1950 - present). From this basic formulation we can assemble higher-level predictions, such as who will get pole position, win a race, or finish on the podium.

This walk-through provides an overview of the model formulation, its parameters, and the approach to tuning those parameters. In general this model attempts to contain the minimum set of parameters and assumptions to generate the highest quality of predictions across the history for Formula One, with a specific set of metrics to ensure high-quality and well-calibrated predictions for drivers and teams at the front of the field.

The system accounts for two types of entities and two components of their performance. The entities are cars (or teams) and drivers. Together a car/driver pair in a single event (race or qualifying session) is referred to as an entrant. For each entity in the data set the system attempts to quantify and predict:

• Reliability: the ability of a car or driver to make it to the end of a given race distance.
• A failure of car reliability is a mechanical failure or some other issue which -- through no fault of the driver -- causes the entrant to not make it to the end of the race distance.
• A failure of driver “reliability” is a crash or other driver mistake which causes the entrant to not make it to the end of the race distance.
• Performance: the speed of the car or driver, separate from their reliability.

It is possible for a car or driver to be incredibly quick but unreliable, or relatively reliable but slow.

Reliability

Reliability is, in effect, an attempt to create a survival model for both car and driver over the course of the race distance: “What is the probability a car or driver fails after X kilometers?”

For example, in a 3 kilometer race with 36 drivers, let us say that:

• in the first kilometer, 4 cars and 2 drivers fail;
• in the second kilometer, 3 cars and 2 drivers fail; and
• in the third kilometer, 3 cars and 2 drivers fail.

	Failure Rate
	Car	Driver
KM 1	4/36 (0.111)	3/36 (0.083)
KM 2	3/29 (0.103)	2/29 (0.069)
KM 3	3/24 (0.125)	2/24 (0.083)

Note that the probability of failure in a given kilometer only considers those entrants still left at the start of that kilometer. There is certainly a possibility of simultaneous failure of both car and driver, but in practice we consider that outcome to be so small as to not be worth quantifying (plus it’s not in the data).

We do not attempt to create a survival model for qualifying sessions for a number of reasons:

1. failure data and reasons are generally not available;
2. the “distance” of qualifying varies widely both over time and over the field, and is not recorded anywhere;
3. some entrants may take only five or six laps to get a qualifying position, while others may take dozens of laps;
4. the final effect of qualifying results on the model is less than that of races in general, so the effect of unexpected failures is overall relatively small.

The question then becomes how to create these survival models. With a few exceptions back in the 50s and 60s, race distances are 300 - 305km. The chart below displays failure rates for the first 300km of race distance, meaning that this data captures the vast majority of mechanical and driver related did-not-finishes (DNFs). Failures are bucketed per 3km in order to de-clutter the chart.

The chart shows:

• the percent of cars still in the race which fail during that bucket (the yellow-boxed scatterplot);
• the percent of drivers still in the race which crash during that bucket (the blue-starred scatterplot); and
• the linear regression of those values and the R2 of the fit (lines).

Two things jump out: the rate of failures are higher for cars than drivers, and the rate of failures is relatively constant throughout the race. Given the small R2 for each regression, this indicates that over a long horizon failure is essentially equally random at any given point in time.

Other analysis of car and driver failures also indicates that there is no interaction between driver and car reliability. In other words, the driver cannot drive in a way which either improves or reduces mechanical failure, at least compared to their peers (with, perhaps, one exception).

If these are true, then we can treat each kilometer traveled as essentially an independent roll of the dice to see if the car fails or the driver crashes. The goal of our model, then, is to predict the odds of failure in an average kilometer for each car and for each driver.

Performance

Performance, contingent on finishing the race, is a proxy for speed. There are several challenges.

• “speed” is a direct combination of both driver skill and car (or team) technical performance;
• there tend to be correlations in that the best drivers are more likely to sign with the best teams, and the best teams have better success at hiring the best drivers; and
• the relative contribution of the car to the overall performance has likely gone up over time, as aerodynamics have played a greater role in overall performance.

Unlike reliability, we can predict raw performance for both races and qualifying. In fact, qualifying has some advantages over races in that everyone competes against everyone else more-or-less equally, and everyone finishes, or at least has a finishing position. This allows us to get a full pairwise comparison of all drivers and cars in a single session.

However, we also take into account three differences between qualifying and races:

• qualifying is shorter in both time and distance, meaning there is less variance. This means that:
  • the same difference in raw speed has more impact on the outcome than in races;
  • there is less information to be gained in qualifying than in a race, so we will exchange fewer Elo points between entrants; and
• the structure of qualifying gives no specific structural advantage to one entrant over another, while races have a starting grid which gives an advantage to drivers at the front.

Our model must attempt to quantify each of those factors.

Other Considerations

Like many other models which evaluate data over sequential events, we incorporate common methods, including:

• regressing per-driver and per-team metrics back to the mean slightly between seasons;
• adding small amounts of uncertainty at the start of each season;
• gradually decaying/aging out old data points over time; and
• limiting the “lookback” window (in addition to aging out the data).

Coming up...

Part II will go into the details of the model implementation. Part III will discuss its predictive performance. Part IV will discuss how predictions get aggregated into metrics which span one or more year.

Hello, analytics community!

 If you've stumbled across this site because of a FiveThirtyEight or Motorsport Magazine article, or a Wharton Moneyball podcast, you've come to the right place.

Analytics is my hobby, and the pandemic robbed me of a bit of my free time. I apologize for the lack of recent content here, but this site is still live and I hope to get at least some semi-regular updates flowing again during this F1 season.

Until then, I'm a bit more active on Twitter and you can find me at @tfgridiron.

Thanks!

Week 17: Saturday Predictions

26 Navy 31  59 Army 26

Key
Close game							Certain victory

Follow us on Twitter at @TFGridiron and @TFGLiveOdds.

Week 17: Top 25 — TFG

Mouse over column headers for definitions, or see this page

Rank	+/-	Team	WinPct	SoS		Adjusted
						Off.		Def.		Pace
1	--	Clemson  ( 12 - 0 )	0.974	0.470	72	28.0	5	7.2	1	163.0	95
2	--	Ohio State  ( 13 - 0 )	0.966	0.661	10	34.2	1	9.7	5	172.9	38
3	+1	Alabama  ( 9 - 2 )	0.934	0.622	16	32.4	2	12.0	14	162.5	98
4	-1	Georgia  ( 10 - 2 )	0.929	0.657	11	21.7	23	8.3	2	159.5	112
5	+3	LSU  ( 12 - 0 )	0.923	0.686	7	32.0	3	12.6	15	171.7	45
6	--	Wisconsin  ( 10 - 3 )	0.920	0.697	6	26.4	6	10.5	8	154.1	126
7	--	Michigan  ( 9 - 3 )	0.912	0.684	9	25.6	7	10.6	9	165.2	81
8	-3	Utah  ( 10 - 2 )	0.902	0.524	59	23.7	13	10.3	6	154.4	125
9	--	Penn State  ( 9 - 2 )	0.899	0.649	13	21.4	27	9.4	4	168.8	61
10	--	Auburn  ( 8 - 3 )	0.884	0.750	1	22.3	19	10.4	7	177.6	16
11	--	Florida  ( 8 - 2 )	0.881	0.655	12	24.6	10	11.6	13	158.8	113
12	+2	Oregon  ( 10 - 2 )	0.876	0.556	46	23.5	14	11.3	11	172.1	44
13	-1	Notre Dame  ( 10 - 2 )	0.866	0.544	54	22.0	20	10.9	10	171.3	47
14	-1	Iowa  ( 9 - 3 )	0.857	0.622	17	17.2	58	8.8	3	160.2	110
15	--	Oklahoma  ( 11 - 1 )	0.809	0.584	34	29.3	4	17.0	57	161.8	101
16	--	Appalachian State  ( 11 - 1 )	0.799	0.384	108	21.6	25	12.8	18	166.7	69
17	--	Washington  ( 6 - 5 )	0.770	0.575	39	20.7	30	13.2	20	164.1	90
18	+7	FL-Atlantic  ( 9 - 3 )	0.765	0.454	76	21.7	24	13.9	25	176.1	21
19	+1	Memphis  ( 11 - 1 )	0.760	0.490	69	24.8	8	16.1	46	173.2	37
20	-2	Baylor  ( 10 - 2 )	0.759	0.554	47	19.6	41	12.7	16	176.4	20
21	-2	Minnesota  ( 9 - 2 )	0.756	0.564	43	24.7	9	16.2	47	156.4	121
22	+1	Boise State  ( 11 - 1 )	0.755	0.401	103	21.9	21	14.3	28	164.8	86
23	-2	UCF  ( 8 - 3 )	0.751	0.426	85	23.9	11	15.8	40	187.5	1
24	-2	Texas A&M  ( 6 - 5 )	0.728	0.701	4	20.1	36	13.9	26	164.9	84
25	-1	Air Force  ( 9 - 2 )	0.711	0.442	82	22.3	18	15.9	43	150.3	130

Rankings through games of 2019-12-08

New entries: none.

Dropped out: none.


Follow us on Twitter at @TFGridiron and @TFGLiveOdds.

Week 17: Full Rankings — TFG

Biggest jumps: FL-Atlantic (0.064); Oregon (0.036); LSU (0.022); Miami-OH (0.021); Boise State (0.018)

Biggest drops: UAB (-0.071); Hawaii (-0.024); Utah (-0.024); Tennessee (-0.021); Georgia (-0.019)

Full rankings after the jump.

Week 17: Top 25 — RBA

Mouse over column headers for definitions, or see this page

Rank	+/-	Team	WinPct	SoS		Adjusted
						Off.		Def.		Pace
1	--	Ohio State  ( 13 - 0 )	0.929	0.536	45	30.3	2	7.2	1	163.8	109
2	--	Wisconsin  ( 10 - 3 )	0.929	0.536	44	23.8	8	7.8	5	160.0	125
3	--	Clemson  ( 12 - 0 )	0.897	0.540	36	25.2	5	7.7	4	169.9	47
4	--	Penn State  ( 9 - 2 )	0.894	0.539	37	22.1	17	7.6	2	165.8	86
5	--	Alabama  ( 9 - 2 )	0.892	0.553	10	28.9	3	10.2	13	158.3	126
6	--	Oregon  ( 10 - 2 )	0.878	0.533	51	23.8	7	8.7	7	178.0	3
7	--	Iowa  ( 9 - 3 )	0.875	0.529	54	19.7	30	7.7	3	163.1	116
8	--	LSU  ( 12 - 0 )	0.870	0.559	7	28.6	4	9.9	11	160.7	121
9	--	Georgia  ( 10 - 2 )	0.858	0.547	28	23.1	10	7.8	6	160.0	124
10	--	Florida  ( 8 - 2 )	0.840	0.562	4	21.2	22	9.8	10	160.4	123
11	--	Notre Dame  ( 10 - 2 )	0.835	0.564	2	22.4	13	10.3	14	165.3	91
12	--	Michigan  ( 9 - 3 )	0.821	0.553	13	22.9	11	10.1	12	164.0	104
13	--	Utah  ( 10 - 2 )	0.819	0.501	66	22.8	12	9.7	9	166.7	80
14	--	Auburn  ( 8 - 3 )	0.816	0.563	3	22.2	16	9.7	8	162.4	119
15	--	Oklahoma  ( 11 - 1 )	0.816	0.543	34	30.5	1	13.7	36	171.8	25
16	--	Iowa State  ( 6 - 5 )	0.795	0.549	22	24.4	6	12.3	23	171.1	30
17	--	Missouri  ( 5 - 6 )	0.773	0.520	59	19.3	32	12.4	25	174.4	11
18	--	UCF  ( 8 - 3 )	0.758	0.471	85	22.2	15	13.7	37	167.3	76
19	--	Washington  ( 6 - 5 )	0.743	0.545	31	21.5	20	12.4	24	167.5	72
20	--	Texas  ( 7 - 5 )	0.720	0.537	42	22.3	14	15.2	50	171.6	26
21	--	TCU  ( 4 - 7 )	0.712	0.503	65	19.8	28	12.0	21	170.8	34
22	--	Duke  ( 4 - 7 )	0.709	0.527	57	21.1	23	15.1	49	169.6	49
23	--	Michigan State  ( 6 - 6 )	0.704	0.553	11	16.1	67	10.9	15	165.7	89
24	--	Cincinnati  ( 10 - 3 )	0.703	0.492	72	18.1	42	11.7	17	170.3	40
25	--	Boise State  ( 11 - 1 )	0.677	0.459	101	19.0	34	13.8	38	169.6	50

Rankings through games of 2019-12-08

New entries: none.

Dropped out: none.


Follow us on Twitter at @TFGridiron and @TFGLiveOdds.

Week 17: Full Rankings — RBA

Biggest jumps:

Biggest drops:

Full rankings after the jump.

Week 16: Saturday Predictions

16 Appalachian State 28  30 LA-Lafayette 25

23 Boise State 36  60 Hawaii 30

82 Central Michigan 29  95 Miami-OH 28

1 Clemson 35  35 Virginia 22

25 FL-Atlantic 30  75 UAB 24

8 LSU 30   3 Georgia 31

20 Memphis 32  31 Cincinnati 28

15 Oklahoma 35  18 Baylor 31

6 Wisconsin 29   2 Ohio State 36

Key
Close game							Certain victory

Follow us on Twitter at @TFGridiron and @TFGLiveOdds.

Week 16: Friday Predictions

14 Oregon 25   5 Utah 29

Key
Close game							Certain victory

Follow us on Twitter at @TFGridiron and @TFGLiveOdds.