With the wraps off the new halo-adorned cars, we launch into 2018. There are several major sources of intrigue at the dawn of the season.
- Can Renault, with three top-budget teams now in their ranks, take the fight to the Mercedes or Ferrari works teams?
- By what margin can 2018 cars exceed the already record-breaking times of 2017?
- With many lower-budget teams inspired by the aerodynamic concepts of the SF70H, how will the midfield reshuffle?
- Can Honda regain some face now that they are allied with Toro Rosso? (The results of this partnership may be inconsequential for the championship, but are vital for the future engine supply of Red Bull.)
In this article, I will first provide an analysis of the times set during testing, focusing on what can be derived from long runs. I will then provide a preview of teammate battles for the season, making predictions using my driver ranking model.
Rationale for long-run analysis
Over the last two years, I have presented analyses of long runs from testing to make forecasts for the season, which have often differed from predictions derived from the headline times. In 2016, I accurately predicted the gaps between the main competitors. In 2017, I focused on the top 3 teams (due to data availability) and accurately predicted that it would be too close to call between Ferrari and Mercedes, while most articles from respected sources were predicting an advantage (0.3-0.8 seconds) to Ferrari.
The reason I have been able to derive relatively accurate estimates is that I tend to ignore the headline times, whereas many media outlets take these times and just add correction factors for tyre compounds. Instead, I perform detailed analysis of long runs performed by the teams, not only correcting for fuel and tyres, but also investigating how much the driver is extracting from the tyres.
Long runs generally provide more insightful data for two reasons.
- First, because it’s harder for teams to disguise their true performance. On shorter runs, we have limited insights into engine modes (race vs. qualifying), how far the driver was pushing the envelope, or how much fuel the car carried. The latter factor alone can account for ~3 seconds per lap — about the difference between Mercedes and Sauber in equivalent conditions! On longer runs, the car has to maintain its performance window to generate representative data. A racing car performs as a package. Detune the engine and you affect some cornering speeds, which changes tyre temperatures, which affects degradation, etc. Perform excessive lifting and coasting and you alter brake temperatures, state of charge, tyre temperature, etc. Due to nonlinear feedback loops, it would be nontrivial to make corrections for these factors, even with access to all the data.
- Second, because we have a long train of lap-times, we can investigate the dynamics of the stint, including how the tyres are degrading.
My methods are certainly not free of uncertainties, especially since I’m working from limited information (lap times and tyre compounds only). Team-based GPS estimates are likely to be more accurate, but teams do have their own agendas when they leak data to journalists, whereas I’m looking at this as an outsider.
My primary data source this year, and in previous years, was @f1debrief on twitter, who posted many of the long runs performed by teams each day of testing. It’s not an exhaustive list of times, but it’s the best resource out there. Due to the abnormal conditions during the first test, I restricted primary analysis to days 5-8 (the second test).
For each stint, I recorded the driver, day of the test, and tyre compounds. I cleaned the data in two ways: (i) lap times more than 1 second slower than the lap time on either side are excluded; and (ii) the first and last lap in each stint is excluded.
For stints with sufficient valid lap times, I fit a linear regression to the valid lap times. Then, to correct for effects of fuel burn, I adjusted the slope by adding 0.054 seconds per lap (my estimate of the cost of one extra lap of fuel, which I derived in 2017 testing).
In stints with low tyre degradation, the linear fit to the cleaned data may yield a negative slope (improving times across the stint), but after fuel correction almost all competitive stints have a positive slope, reflecting performance loss from the tyres as they degrade. The value of the fuel-corrected slope is largely dependent on two factors: (i) the tyre compound (softer compounds degrade more rapidly), and (ii) how hard the driver is pushing (tyres degrade more rapidly when the driver is pushing hard).
Plotting all of the stints for all the drivers across the second test, we see the following relationship.
In general, stints with higher tyre degradation have lower intercepts. This is entirely expected, since a driver who is pushing harder will have lower times (at least at the beginning of the stint, before the tyres fall apart). It’s also consistent with what was found in the past two seasons.
In the above plot, we can see a significant amount of scatter. Even for points of the same color with approximately the same x-value (Fuel-corrected slope), we see a range of y-values (Intercept). This is because we are not correcting for other factors that significantly affect the intercept. These may include: differences in fuel load, differences in car performance between teams, and other factors such as engine modes, set-up, and track conditions. We therefore also need to correct the intercept to make useful inferences.
For a given team, we can create reference points using confirmed race simulations (i.e., connected stints for which the fuel load is known, since the team only changed tyres at intervening stops). For other stints, we can estimate fuel loads using the following information: (i) the car could not have less laps of fuel than the length of the stint, or more than a full tank of fuel, which generates constraints; and (ii) for teams with race simulations, we can find best fits for unknown stints to reference estimates, given plausible fuel levels at which the stint could have been run.
Using this approach, I fuel corrected all the representative long runs (throwing out stints with very few valid laps or abnormally slow times). Below is a graph showing the results for the medium compound, which was the most used compound during the test, and the backbone of most race simulations.
The picture at the top of the hierarchy is clear. Only three teams came out of this test looking capable of challenging for wins (and perhaps even podiums). Of those, Mercedes is clearly the current favorite, with an estimated edge of half a second over the nearest opposition. The ordering of Ferrari and Red Bull is unclear, but a narrow edge to Ferrari emerges here on the medium compound.
A full second back from Ferrari and Red Bull, we have Renault, and perhaps McLaren, as best of the rest. Compared to other top teams, McLaren were unusual in shying away from the medium and soft compounds during testing, doing more running on the supersoft, ultrasoft, and hypersoft compounds. The one representative stint on mediums occurred during a race simulation by Vandoorne, and it would suggest McLaren have pace comparable to Renault, around 1.0-1.4 seconds slower than Ferrari.
No Force India race simulations were posted online, but I would estimate their position to be a few tenths behind Renault, based on the long runs they delivered.
The competitive position of Mercedes is determined not only from their race simulations, but also one very impressive run, shown below.
On day 6, Bottas unleashed an incredible 8-lap stint. Every time posted was in the 1:19s. The most conservative possible interpretation is that Bottas started this stint with only 8 laps of fuel on board. That assumption actually corresponds to the plotted position of this stint above. If Bottas had any more than 8 laps of fuel on board, the point should actually be plotted further below.
Since not all teams were well represented in the times posted for mediums, we can gain additional insights by studying long runs on softer compounds. Using my best estimates at fuel loads, the fuel-corrected stints are shown below.
Only three of the soft stints were from confirmed race simulations (plotted as squares), making direct comparisons more difficult than on the medium compound. Again, Mercedes have a clear edge on anyone, but only around 0.3-0.5 seconds ahead of Ferrari. This is consistent with the report that Mercedes were relatively struggling with the soft compound compared to mediums.
Renault’s placement, around 1.3 seconds behind Ferrari, is very consistent with the gap I observed on mediums. On softs, Haas looked surprisingly competitive, just over a second behind Mercedes, and probably a few tenths ahead of Renault, rendering them 4th quickest.
McLaren’s only long run on softs was during Vandoorne’s race simulation, and seemed abnormally slow compared to other McLaren stints, suggesting another factor at play, or perhaps an inaccurate fuel correction.
Williams looked quite slow on their stint with Stroll driving, around 2.5-3.0 seconds slower than Mercedes at an equivalent tyre degradation rate.
Finally, a few teams ran race simulations on the supersoft tyre, usually to simulate the first race stint on heavy fuel loads. McLaren in particular favored running the supersoft compound throughout the test. On the basis of these data points, we see McLaren trailing Ferrari by around 1 second on average, consistent with the margin on the medium compound race simulations.
On the supersoft, Haas appear to be a few tenths behind McLaren. It will be interesting to see in Melbourne whether they slot in ahead or behind the McLaren/Renault cars, as these data remain open to interpretation.
Notably, I could not find any sufficiently informative stints for Toro Rosso or Sauber to establish their positions in the hierarchy. Mark Hughes’ interesting analysis suggests that Sauber are last (I don’t disagree, from the times I have seen), and that Toro Rosso-Honda have an advantage over the struggling Williams and even Force India. The main positive for Honda at this stage is reliability — something they lacked entirely over the past three preseasons.
So, what about Ferrari’s headline times? Vettel smashed the Barcelona track record with a 1:17.1 on the hypersoft compound, and Raikkonen’s 1:17.2 on the final day showed such times could be achieved with either driver at the wheel. Does this not indicate that Ferrari are now top dogs?
I don’t believe so. If we consider that Bottas achieved a 1:19.38 on mediums, and furthermore consider Pirelli’s estimated time difference of ~2.5 seconds between fresh mediums and fresh hypersofts, then this translates into an achievable time of 1:16.9 on hypersofts. The way in which Mercedes achieved this time is particularly ominous and unlike comparable times set by other teams. Bottas’ 1:19.38 was in the middle of the 8-lap stint (plotted above) where all 8 laps were within a close range (1:19.38 to 1:19.97).
- The car was not in a one-lap qualifying mode (typically worth a few tenths).
- Nor was the car running on fumes (it could have been at least 3 laps lighter, which is worth about another 0.15 seconds)
- The tyres were not degrading rapidly, implying they could stand additional punishment
With those factors in mind, a lap time in the low 1:16s seems achievable for Mercedes on a glory run on hypersofts. Ferrari were probably still well within their car’s limits on their best hypersoft laps, but are they able to go that quick? Unless they are sandbagging in a fashion unlike any recent preseasons, the data suggest not.
In summary, looking at long runs and individual lap times, I think we can be quite confident that Mercedes have a lead on the field at this moment. Short of an upset, or a major development, Ferrari and Red Bull will be filling out the podium at most races. While I personally hoped to see the works Renault and McLaren teams entering the fray, I think the gap back to the fourth-best team remains a significant divide, at least for now. The silver lining is that Ferrari underwent a major redesign this year, while McLaren have a new engine partner and are clearly running in a compromised state with poor reliability. Both teams therefore have significant potential for rapid improvement within the season, whereas Mercedes are refining an already excellent concept.
Using my model of driver and team performance across F1 history, I made predictions of the likelihood of each driver outperforming their teammate in 2018. “Outperforming” is defined as scoring more points per counting race across the season (i.e., excluding non-driver DNFs, such as mechanical failures). The current version of the model takes age and experience effects into account, with important impacts on predictions in some match-ups, noted below.
There are no predictions for Toro Rosso, Williams, or Sauber, due to drivers for these teams having insufficient counting races to date.
At Mercedes, Hamilton is predicted to retain the advantage over his teammate, although a 1 in 4 chance of Bottas winning can’t be totally discounted. Given the way the Mercedes car is looking after the tests, Hamilton is looking good for a fifth title.
At this stage in his career, Raikkonen is clearly playing a purely supportive role to Vettel. If the Finn were to come out on top in 2018, it would be a major upset and contrary to the form guide of the past several years. In 2017, there was only one grand prix where Raikkonen genuinely outpaced his teammate. With Raikkonen turning 40 next year, that should perhaps be no major surprise. Some will wonder, how might prime Raikkonen have fared against prime Vettel? We can simulate this by correcting for both age and experience effects. In that hypothetical instance, the model gives a 74% advantage to prime Vettel over prime Raikkonen.
After the fireworks of last year, this should be an intense teammate battle. While Perez consistently had the edge on Ocon in the first half of 2017, Ocon seemed to have the measure of Perez in the second half of the season. Given the benefit of additional experience and age, the model sees Ocon as a narrow favorite going into 2018. Perez won’t want to let that happen, so it should be an interesting fight, even if it’s no longer for potential podium places or even points, given the number of competitors who presently look quicker than Force India’s 2018 offering.
As we saw at the end of last season, the model is quite enamored of Sainz at the moment. As I explained in that analysis, this has arisen partly due to the lack of outward teammate connections from the Red Bull driver cluster. Currently, the model has limited ability to compare Sainz to drivers other than Ricciardo, Verstappen, and Kvyat. I therefore view this prediction with a healthy dose of skepticism. The Sainz-Hulkenberg match-up at Renault late last year was unfortunately quite uninformative due to Renault’s atrocious reliability. I’m eagerly looking forward to the data generated by 2018. On paper, it looks like they will be sharing track with McLaren, which could lead to some suberb battles.
If Haas have achieved a meaningful step forward, as testing times suggest, this could be a teammate battle with increased screen time. On the basis of last year, we would expect the battle to be close. The model sees it as close to a 50-50, but interestingly with a slight edge to Magnussen. Last year, he was in only his third full season of F1, and with a career interruption in 2015, meaning he still has some scope for improvement.
On that note, I’ll see you in Melbourne!