Design one metric for your weakest phase
Generated from
content/lms/data-interpretation-ii-advanced/06-metric-driven-analysis/03-custom-metrics.md; edit the source file, not this page.
Source path: content/lms/data-interpretation-ii-advanced/06-metric-driven-analysis/03-custom-metrics.md
Course: Read the data your hands can't feel
Module: Build metrics that survive the noise
Estimated duration: 45 minutes
What you are building
A personal metric is not another page of data. It is a small instrument you design for one driving problem, in one phase of the lap, so the next session has a clear job. The normal reports tell you where to look. Segment and section times, fastest rolling, theoretical fastest, throttle traces, brake-pressure traces, steering, RPM, gear, G-sum, and GPS line can all point toward a loss. Your custom metric turns that pile of evidence into one repeatable question: did the behavior I chose get better, worse, or stay the same?
That is the core distinction. A report describes the lap. A personal metric directs your practice. If your data review ends with a general statement like be smoother, brake better, or get on throttle earlier, you do not yet have a metric. You have a wish. A metric names a behavior that can be seen in the data, checked against other channels, compared against a better reference, calibrated to what you felt in the car, and converted into one objective for the next session.
The rule: one phase, one behavior, one objective
Build the metric around one weak phase, not the whole lap. The bonded data-process material gives you the sequence: start with the overview, look for incongruencies, dig for details, use other channels when they are available, ask why, compare when you can, calibrate to your driving, imagine the ideal, and set objectives for the next session. That process is the spine of a good custom metric.
The mistake is trying to make the metric impressive. You are not designing a championship engineering dashboard. You are trying to make your own driving easier to improve. Keep it simple and focus on the basics. If the metric cannot survive a quick review between sessions, it is probably too complicated for driver development. You should be able to explain it in one sentence before you look at the traces, and you should be able to score it the same way after every session.
A useful personal metric has five parts. First, it names the phase you are attacking: brake phase, release and entry, middle of the corner, exit throttle, fast-corner commitment, shift or gear choice, or another narrow section your data actually supports. Second, it names the behavior, such as coasting, hesitant throttle application, early throttle followed by a lift, long brake-pressure tail, inconsistent brake pressure, light-and-long braking, hard-and-short braking, or lift in a fast corner. Third, it names the channel that will show the behavior. Fourth, it names at least one cross-check channel so you do not fool yourself. Fifth, it names the objective for the next session.
Choose the weak phase without redoing the segment lesson
This module already has segment-report lessons. Do not duplicate that work here. Use the segment report as the doorway, then move on. If one section is slow, inconsistent lap to lap, or far away from your fastest rolling or theoretical-fastest possibility, that section becomes the candidate. The custom-metric question is not simply which section is slow. It is why that section is slow in your driving.
Start with the overview. Look across the lap before you zoom in. If every loss is random, you may not have a single weakest phase yet. If the same section keeps appearing, or if one phase keeps showing the same trace shape, you have a better target. A throttle trace that shows repeated coasting, hesitation, or early application followed by a lift gives you one kind of metric. A brake-pressure trace with an inconsistent shape, a long tail, or a light-and-long pattern gives you another. A fast corner with repeated lifts points to a different metric again.
The phase should be small enough that you can drive with it in mind. Whole-lap metrics tend to become vague. Whole-corner metrics can still be too broad if the corner contains braking, release, rotation, maintenance throttle, and exit. The stronger choice is a phase you can recognize while driving and verify afterward. For example, the target might be the throttle behavior in one exit segment, the brake-pressure shape in one braking zone, or the lift behavior in one fast-corner section.
Translate the phase into a behavior
A weak phase becomes trainable only when you turn it into behavior. Braking better is not a metric. Brake-pressure repeatability in the target zone is closer. Being faster on exit is not a metric. Clean exit throttle in the target section is closer. A clean metric should point to something the trace can reveal and something you can practice in the car.
For throttle-phase work, the bonded data list gives you several behavior choices: coasting, hesitant application, early application leading to a lift, and lifts in fast corners. These are better than a vague throttle score because they describe different problems. Coasting may mean you are waiting between controls. Hesitant application may mean you are unsure whether the car is ready. Early application followed by a lift may mean the first throttle request was not compatible with the line, grip, or commitment you had. A lift in a fast corner may mean the trace is exposing a confidence or placement problem. The metric should distinguish these instead of lumping them together.
For brake-phase work, the bonded data list points at the brake-pressure trace shape: initial application, trail, long tail, inconsistent pressure, and the contrast between light-and-long and hard-and-short braking. Again, these are not the same problem. A long tail is different from inconsistent pressure. Light-and-long braking is different from a clear initial application and a controlled release. If your metric treats all braking loss as one number, it may hide the exact habit you need to change.
For supporting channels, the corpus names steering, RPM, gear, segment and section reports, fastest rolling, theoretical fastest, G-sum, GPS line, total steer angle, and throttle histogram. You do not need every channel every time. You do need enough cross-checking to avoid blaming the wrong thing. If the throttle trace looks clean but acceleration rate is poor, check gear and RPM before deciding the driver was timid. If the brake trace changes but the GPS line changes too, the braking metric may be mixed with a line change. If a fast-corner lift appears, total steer angle and GPS line may help you decide whether the lift followed a placement problem or a commitment problem.
Define the metric so it can be scored after every session
A personal metric needs a scoring rule. The rule can be simple. It can be a count, a duration, a yes-or-no classification, or a clean-lap percentage. What matters is that you score the same behavior the same way every time.
A count works when the error is discrete. Count how many laps in the target section show early throttle followed by a lift. Count how many laps show a lift in the fast corner. Count how many target braking zones show a long brake-pressure tail. Counts are easy to use between sessions, and they match the keep-it-simple rule.
A duration works when the error is waiting. Measure the amount of coasting in the target section, or the hesitation between your intended throttle application and the point where the trace actually shows a confident application. If your software does not make that easy, do not turn the review into a spreadsheet project. Use a count or a simple short-medium-long classification. Getting your hands dirty with the data matters more than pretending the first version of the metric is laboratory precise.
A classification works when shape matters more than exact size. For brake pressure, score each target lap as clean initial-and-release shape, long tail, inconsistent pressure, or light-and-long. For throttle, score each target lap as clean application, coast, hesitant application, early-then-lift, or fast-corner lift. This gives you a pattern you can actually discuss with a coach or compare to a better lap.
A percentage works after you have a reliable classification. If seven of ten target laps are clean exits, your clean-exit percentage is 70 percent. If three of ten laps show a fast-corner lift, your fast-corner lift rate is 30 percent. The percentage is not magic. It is just a compact way to watch whether the chosen behavior is becoming more repeatable.
Cross-check before you believe it
A custom metric becomes dangerous when it flatters a single channel. The corpus is explicit about using other channels when available and looking for incongruencies. Incongruency is not failure; it is a clue. If the primary channel and the supporting channels disagree, the metric has found the next question.
Suppose your exit metric says throttle hesitation improved, but the segment time did not. That does not automatically mean the metric failed. It may mean the line changed, the gear changed, the car was slower earlier in the section, or the improvement is not yet big enough to show in the section time. Check RPM, gear, GPS line, and the segment report. Ask why before you throw away the metric.
Suppose your brake-shape metric looks better, but the car arrives at the next phase worse. Check whether the brake-pressure improvement came with a different steering trace or line. A cleaner-looking brake trace that compromises the next phase may not be an improvement. The point of the metric is not to make one trace pretty. The point is to improve the driving phase without creating a new problem in the rest of the section.
Suppose your fast-corner metric says you stopped lifting. That is good only if the rest of the evidence agrees. Check the GPS line and G-sum if you have them. Check the section time. Check your memory of the lap. Calibrate to your driving. If the data says the lift disappeared but you remember adding a big steering correction, keep digging. The metric is a prompt to investigate, not permission to stop thinking.
Use comparison, but do not worship the reference
Comparison is useful when you have it. Compare to your own best lap, your fastest rolling possibility, your theoretical fastest, a coach lap, or another driver only when the comparison is fair enough to teach you something. The corpus supports comparing when you can, not forcing comparison when you cannot. If all you have is your own data, compare your weak phase to your best version of that same phase.
The ideal trace is also useful, but only as a sketch. The process material says to imagine what ideal would look like. That means you should be able to describe the trace you are trying to create: no coast in the target exit, no early throttle followed by a lift, a brake-pressure shape with the intended initial application and release instead of a long uncertain tail, or a fast-corner section without an unnecessary lift. The imagined ideal gives direction. It is not evidence by itself.
This matters because theoretical fastest and fastest rolling can tempt you into chasing a fantasy lap. They are excellent for locating opportunity, but your custom metric should remain attached to one behavior you can practice. If the fastest rolling report says a section has potential, the next question is still practical: what behavior in that section can I change next session, and what channel will prove whether I changed it?
Calibrate the metric to what you felt
Data work improves fastest when you connect it to the driving memory. The process material says to calibrate to your driving. That is a driver skill, not an engineering luxury. After the session, ask what the weak phase felt like before you look at the trace. Did you feel the coast? Did you know you hesitated? Did you sense the early throttle that led to a lift? Did the brake release feel long, vague, or inconsistent? Then check the data.
If your memory and the trace agree, you have a strong learning loop. You can start recognizing the behavior in the car. If your memory and the trace disagree, do not be embarrassed. That is exactly why you are using data. The metric can teach you what the behavior feels like. Over time, the goal is not just to make the trace better. The goal is to become a driver who can predict what the trace will show.
This is also where a coach or instructor can help without turning the metric into a lecture. Bring one clear metric and one clear question. Instead of asking what am I doing wrong everywhere, ask whether the target exit shows coasting, hesitation, or early throttle followed by lift. Instead of asking whether your braking is good, ask whether the trace shape shows the intended initial application and release or a long tail. A specific question produces a specific answer.
Set the next-session objective
A metric is unfinished until it produces a next-session objective. The corpus process ends with setting objectives for the next session, and that is where many drivers stop short. They build a clever metric, admire it, and then go back on track with no driving instruction for themselves.
A good objective is small and behavior-based. In the target exit, reduce early-throttle-then-lift events. In the target braking zone, make the brake-pressure shape more repeatable. In the target fast corner, remove unnecessary lifts while checking that line and steering do not become messy. These objectives are not grand. That is why they work. They give you one job you can remember at speed.
Do not change three things at once. If you change brake timing, line, gear, and throttle commitment in the same session, your custom metric loses clarity. You may go faster, but you will not know why. You may go slower, and you will not know what to fix. Keep the objective narrow enough that the data can answer it after the session.
What improvement looks like
Improvement is not always immediate lap time. Sometimes the first win is that the trace becomes easier to read. The exit metric shows fewer coast or hesitation flags. The brake metric shows fewer long-tail laps. The fast-corner metric shows fewer lifts. The section time becomes more consistent lap to lap. The supporting channels stop contradicting the primary channel. Your memory of the lap becomes closer to the data.
Lap time still matters. Segment and section reports exist for a reason. But in a metric-driven practice session, lap time is the result, not the only measurement. If your behavior improves and the section time follows, you have a strong signal. If the behavior improves but section time does not, dig for details. If section time improves but the behavior did not, ask whether another variable changed. The purpose is to learn why the lap changed, not simply whether it changed.
Keep learning, and know when the metric points beyond driving technique
Some findings are bigger than your current metric. The driving-data corpus points you toward continued learning and asking why. The chassis and suspension sources also remind you that understanding what the car is doing is part of the driver's job. If the data suggests that a driver input is being shaped by car behavior, do not hide that inside a driver-only metric. You may need to ask someone to explain the setup, suspension, or vehicle-dynamics side. The metric can still be useful, but it should not pretend that every pattern is purely a driver habit.
That is why the best personal metric is humble. It does not claim to explain the entire lap. It gives you one disciplined way to work: overview, weak phase, behavior, channel, cross-check, comparison, calibration, imagined ideal, next-session objective. Use it for a session, review it honestly, and either keep it, refine it, or retire it when the weak phase moves somewhere else.
Worked example: clean-exit percentage for throttle hesitation
Your segment report shows that one exit section is costing time and that the loss changes from lap to lap. You open the throttle trace for that section and look for the behaviors the corpus names: coasting, hesitant application, early application leading to a lift, and any lift in a fast corner if this section is fast enough for that to matter.
Do not begin with a complicated throttle score. Start with a simple classification for each valid lap in the target section. Mark the lap clean if the throttle application matches the behavior you intended and does not show the problem you selected. Mark coast if the trace shows a waiting period. Mark hesitant if the application is uncertain rather than committed. Mark early-then-lift if the throttle comes in and then has to come back out. If the section is a fast corner, mark lift if the trace shows the lift you are trying to remove.
The personal metric is clean-exit percentage. If ten laps are valid and six are clean, the score is 60 percent. The supporting review is just as important as the percentage. Check segment time to see whether the cleaner exits matter. Check RPM and gear so you do not confuse a gear-choice issue with a throttle issue. Check GPS line if available so a changed path does not masquerade as better throttle work. If acceleration rate remains weak while the throttle metric improves, dig for details instead of declaring victory.
The next-session objective is specific: raise the clean-exit percentage in that same section without creating early-then-lift events. The success cue in the car is that you know what a clean application feels like before you see the graph. The success cue in the data is fewer coast, hesitant, and early-then-lift classifications, with the section time becoming more consistent or faster rather than merely different.
Worked example: brake-shape repeatability for a target braking zone
Your overview shows one braking section where the section time is inconsistent. You open the brake-pressure trace and do not ask whether the braking is good in general. You ask what shape appears in this one zone. The corpus points you toward initial application, trail, long tail, inconsistent pressure, and the contrast between light-and-long and hard-and-short braking.
Build the metric as a shape classification. For each valid lap through the target braking zone, mark the trace as intended shape, long tail, inconsistent pressure, light-and-long, or hard-and-short. The intended shape depends on your plan for that corner and car, so do not invent a universal ideal. What matters is whether the actual trace matches the behavior you are trying to repeat and whether the other channels support it.
Now cross-check. Use the segment or section report to see whether the shape classification relates to time. Use steering or total steer angle if available to see whether brake release and steering demand are tangled together. Use GPS line to check whether an apparent brake improvement came from a different path. Use speed or acceleration information if your system gives it, because a better-looking brake trace that leaves the car worse for the next phase is not a real improvement.
The next-session objective is to reduce the unwanted shape, not to chase an abstract braking number. If the weak pattern is a long tail, work on making the release match your intended shape in the target zone. If the weak pattern is inconsistent pressure, work on repeatability first. If the weak pattern is light-and-long or hard-and-short, use the metric to discover which version costs the section. Success is a more repeatable classification pattern, fewer incongruencies with other channels, and a clearer connection between what you felt in the braking zone and what the trace later shows.
Common mistakes
The first mistake is building a vanity metric. A vanity metric looks clever but does not change your next session. Whole-lap time, theoretical fastest, or a large composite score can be useful context, but they are too broad if they do not tell you what to do with your hands and feet in one phase. Good looks like one behavior in one section with one next-session objective.
The second mistake is making the metric too complicated. If you cannot score it between sessions, you will not use it. Good looks like a count, duration, classification, or percentage that survives real paddock time. Keep it simple, focus on the basics, and get your hands dirty with the data before inventing a more elaborate number.
The third mistake is trusting one channel alone. A throttle trace can identify coasting, hesitation, early application followed by lift, or fast-corner lift, but it may not explain why the behavior happened. A brake trace can show shape, trail, long tail, inconsistency, or a light-and-long pattern, but it may not prove the whole corner improved. Good looks like a primary channel plus at least one supporting channel such as segment time, steering, RPM, gear, G-sum, GPS line, total steer angle, or throttle histogram when those channels are available.
The fourth mistake is confusing comparison with proof. Fastest rolling, theoretical fastest, and other-driver comparisons can show opportunity, but they do not automatically define the driver action. Good looks like using comparison to select the target, then designing a behavior metric you can actually practice.
The fifth mistake is ignoring incongruencies. If the metric improves but the section report does not, or if the trace improves while another channel gets worse, do not throw away the data and do not force a happy story. Good looks like asking why and digging for details.
The sixth mistake is failing to calibrate the metric to your driving. If you only inspect the graph after the session, the lesson stays on the laptop. Good looks like predicting what the trace will show, checking whether your memory matches the data, and using the mismatch to sharpen your feel.
The seventh mistake is changing the objective too soon. If you pick a new metric every session, you may never learn whether the first one worked. Good looks like holding one metric long enough to see whether the chosen behavior changes, then using the trend and the next segment review to decide whether to keep, refine, or retire it.
Drill: the three-session weak-phase metric loop
Run this drill over three consecutive sessions at your next event. Use only one target section and one metric. If your event sessions are about twenty minutes, spend the whole session driving normally, but review only the target section afterward. If your sessions are shorter or longer, keep the structure and adjust the review time.
Before session one, choose the candidate section from your segment or section report. Write one sentence that names the phase and the suspected behavior. Examples are target exit throttle hesitation, target braking-zone long tail, or fast-corner lift rate. Do not set a correction yet. Session one is the baseline. After the session, score every valid lap in that section using your simple rule. Then check at least two supporting pieces of evidence, such as segment time plus RPM and gear, or brake trace plus GPS line, or throttle trace plus total steer angle.
Before session two, set one objective from the baseline. The objective must be behavior-based. Reduce early-throttle-then-lift events. Reduce long-tail brake classifications. Reduce fast-corner lift count. Drive the session with that single objective. Afterward, score the same metric the same way. Do not add a new metric just because another graph looks interesting.
Before session three, decide whether the metric is clear enough to repeat. If session two improved the behavior and the supporting channels did not reveal a new problem, repeat the same objective and look for consistency. If session two produced incongruencies, keep the same metric but change the question to why. For example, if clean-exit percentage improved but section time did not, check line, gear, and RPM before deciding the throttle objective failed.
Call the drill successful when three things happen at once: the chosen behavior improves by your own scoring rule, at least one supporting channel does not contradict the improvement, and your memory of the phase is closer to the data than it was in session one. That last point matters. You are not only building a number. You are training yourself to feel the behavior the number exposes.
When this principle breaks down
The custom-metric method breaks down when the data bond is too thin. If the channel you need is not available, do not pretend. Use the channels you have, but reduce the confidence of the conclusion. A throttle-hesitation metric is stronger with throttle, segment time, RPM, gear, and line than it is with throttle alone. A brake-shape metric is stronger with brake pressure and supporting line or steering information than it is with a single trace viewed in isolation.
It also breaks down when the behavior is not actually the driver habit you think it is. If your review keeps pointing toward setup, suspension behavior, sensor accuracy, or vehicle-dynamics questions you do not understand, stop hiding that behind a driver metric. The driver still has to learn what the car is doing. Ask someone who can explain it, return to the relevant books or resources, and keep the metric humble until you understand the mechanism.
Finally, the principle breaks down when you use the metric as a scoreboard instead of a learning loop. The process is overview, incongruencies, details, other channels, why, comparison, calibration, ideal, and next objective. If you skip the why, skip the calibration, or skip the next objective, you no longer have a driver-development metric. You only have another graph.
Author Review
No quiz questions are attached to this lesson.
Sources
| # | Document | Chunk | Pages | Score | Collection |
|---|---|---|---|---|---|
| 1 | Data for Drivers | cabda699642b26311b0a7ef998da2c71 | 15 | 1 | uio_books_raw_v1 |
| 2 | Data-for-Drivers-PRINT | bbb02386-778f-20ec-ad16-b9c016921743 | 16 | 1 | uio_books_raw_v1 |
| 3 | Data-for-Drivers-PRINT | b80dc634-a0a7-d6de-d470-353aed47e2a6 | 17 | 1 | uio_books_raw_v1 |
| 4 | Data for Drivers | 27ec1aea-60bb-f052-9a1a-294b72597f55 | 17 | 1 | uio_books_raw_v1 |
| 5 | Analysis Techniques for Racecar Data Acquisition | 74ff9a5e-1293-62fd-0a20-d45fbbd732b7 | 4 | 1 | uio_books_raw_v1 |
| 6 | Racing Chassis and Suspension Design Carroll Smith | 52047a73-bbbf-e4e8-51ff-bb6cdbc0101b | 134 | 1 | uio_books_raw_v1 |
| 7 | Ultimate Speed Secrets - Ross Bentley | 149c4d5c-d228-0358-acc0-8a92ac07ec7c | 50 | 1 | uio_books_raw_v1 |
| 8 | Speed Secrets Professional Race Driving Techniques Ross Bentley | 26bc8e35-76a6-4f72-ea86-df10ba43a636 | 14 | 1 | uio_books_raw_v1 |