1. Start with decisions, not an algorithm

These are potential uses inferred from the analysis, not outcomes tested on a business, audit engagement or national registry. Where a valid measured rating is available, use it rather than replacing it with a model estimate.

The question

Can we estimate a vehicle's reported tailpipe CO₂ rating, and explain how that rating can inform decisions? We examine two information settings: specifications only, and fuel consumption already known. Those settings solve different practical problems.

The unit

The target is g/km: grams of CO₂ per kilometre. It is an emissions intensity, not annual tonnes. An annual estimate needs a distance. It is also different from atmospheric CO₂ concentration (ppm), all greenhouse gases (CO₂e), and lifecycle emissions.

2. Understand linear regression first

For one predictor, a fitted model has the form:

With two predictors:

Ordinary least squares chooses coefficients to minimise the sum of squared residuals. A residual here is actual − predicted. Coefficients describe the fitted association, not an independent causal effect. Engine size and cylinders are strongly related, so their separate coefficients require cautious interpretation.

Ridge adds a penalty on coefficient size. Random forests fit multiple decision trees. We compare them with simple alternatives; a more complex model must earn its place through validation.

What changed from the supplied v3 notebook?

The v3 notebook provides more explanations than the first upload, but it fits X and y after creating a split, then evaluates predictions on the same full dataset. Its RMSE ≈29.96 and R² ≈0.738 are in-sample results. Its fitted best-fit line is also described as an ideal 1:1 line, and one example uses g/kg instead of g/km. This edition fits training partitions only, reports held-out performance, and uses a true identity line for the prediction plot.

3. Set up a reproducible notebook

Place CO2 Emissions_Canada.csv beside the notebook. In Jupyter, choose Restart Kernel and Run All Cells. The fallback upload/ location is useful when working with the attached files. The package includes requirements.txt with the analysis versions used for this run.

Imports are split by role. Shared helpers only format outputs or construct reusable pipelines; the individual analysis steps remain visible below.

Step 1. Import data-handling tools

These tools load tables, create arrays and record the input fingerprint.

Step 2. Import plotting tools

Matplotlib draws figures; seaborn provides readable statistical plots.

Step 3. Import model tools

Each estimator below will later be created in its own cell.

Step 4. Import preprocessing tools

Pipelines fit these transformations inside the appropriate training partition.

Step 5. Import validation tools

Grouped splitting keeps records of one exact vehicle model together.

Step 6. Choose the input file

Only the supplied CSV is used. No replacement dataset is downloaded.

Step 7. Record provenance

The exact source link comes from the supplied v3 notebook. Its licence was not verified.

Step 8. Define the analysis settings

The test share is a share of model groups, not an exact share of rows.

Step 9. Set readable chart colours

The palette matches the portfolio; chart colours still retain a clear legend.

Step 10. Create the interpretation helper

Jupyter can display Markdown. This helper keeps interpretation cells short and editable.

Step 11. Create the figure helper

One helper handles layout and optional export; each chart gets its own analysis cell.

4. Inspect the data before making claims

Rows are catalogue/rating entries. They are not vehicles sold, and their frequency is not market share. The file does not include a model year, a sales quantity, annual distance, VIN, or tax certificate identifier. Do not assume that these entries represent all vehicles on a country's roads.

The source names fuel consumption in L/100 km and mpg. We use L/100 km consistently and avoid treating an inverse unit conversion as an independent physical measurement.

Step 12. Load the raw CSV

Keep a raw copy so every later transformation can be traced.

Step 13. Count rows and columns

Start by explaining what one row represents.

Step 14. Preview five records

A small preview is more readable on a website than printing the entire dataset.

Step 15. Check data types

Numeric columns should be numeric before correlations and model fitting.

Step 16. Check missing values once

isna() and isnull() are aliases; repeating both adds no new evidence.

Step 17. Count exact repeated rows

Repeated rows can influence both descriptive summaries and model fitting.

Step 18. Define shorter column names

The source names and units remain documented; shorter names make learning the code easier.

Step 19. Check the expected schema

Stop if a different CSV has been supplied instead of silently using the wrong columns.

Step 20. Trim text without discarding records

Leading or trailing spaces can create false differences between model identities.

Step 21. Check numeric validity

Negative or zero combustion-vehicle inputs require investigation before fitting.

Step 22. Count fuel categories

Sample size matters when interpreting averages or model reliability.

Step 23. Check the combined-rating arithmetic

The documented combined formula is 55% city and 45% highway.

Step 24. Summarise the target

Use the distribution to give later prediction errors a meaningful scale.

5. Protect the test set

An ordinary random row split may put repeated model entries on both sides. We instead hold out complete combinations of Make + exact Model text. This is an imperfect but useful separation: related trims and model families can still have different names. It does not simulate the next year of sales, because a verified year is unavailable.

The following cells split the data before model exploration. All exploratory plots use training records. The main selection uses grouped cross-validation inside that training partition.

Step 25. Create the grouping key

The two-column identity avoids collisions caused by simple string concatenation.

Step 26. Create a grouped hold-out split

A fixed seed makes the same group selection reproducible.

Step 27. Store training and test records

Use copies so later plotting or audit columns do not alter the input unexpectedly.

Step 28. Verify the separation

A useful check tests the intended boundary directly.

Step 29. Inspect fuel coverage across partitions

A model cannot claim validated category performance when the category is absent from testing.

6. Explore one relationship at a time

A useful chart answers a question. We use targeted plots rather than a large, difficult-to-read pair plot. Association is not proof of causation, and a visual pattern is not a statistical test.

Step 30. Plot the distribution of training emissions

This shows where most ratings lie and whether a few large values extend the range.

Step 31. Calculate a correlation matrix

Pearson r measures the strength and direction of a linear association.

Step 32. Plot a readable heatmap

Red is negative and blue is positive. A large negative correlation is still strong.

Step 33. Plot engine size against emissions

This is the relationship used by the simplest regression model.

Step 34. Plot fuel consumption by fuel type

Separate colours reveal whether a common fuel-consumption slope misses structure.

Step 35. Summarise vehicle classes in training

Class means describe this catalogue; show sample counts alongside them.

Step 36. Plot the most frequent classes

Boxplots show the spread within a class, which a mean alone would hide.

7. Learn fitting with two small linear models

We first fit engine-only and engine-plus-cylinders models on training records. These are declared learning comparators, not models chosen after inspecting test errors. Both join the validation comparison later.

The fitting and prediction steps have separate cells so their purpose is explicit. No full-data fit is used to report test performance.

Step 37. Choose the engine-only training inputs

X is a table of predictors; y is the target series.

Step 38. Create an engine-only estimator

Creating a model does not fit it yet.

Step 39. Fit the engine-only model

The estimator learns coefficients from training observations only.

Step 40. Read the engine-only coefficients

The slope's unit is g/km for one additional litre of engine size.

Step 41. Create one engine-only prediction

Use a DataFrame with the same feature name used during fitting.

Step 42. Choose the two-predictor training inputs

Two related predictors can describe more of the observed variation.

Step 43. Create a two-predictor estimator

This model keeps coefficients in the original units.

Step 44. Fit the two-predictor model

Use the same training records as the other models.

Step 45. Read both coefficients

Each coefficient holds the other predictor fixed in the fitted model.

Step 46. Predict for a two-predictor example

An estimate from two specifications is not a tax certificate.

8. Compare models fairly

All candidate choices below are fixed before final test reporting. Specifications-only and fuel-known models use different information; their difference is not solely an algorithm effect. Scaling is helpful for Ridge's penalty, and all fitted transformations stay inside pipelines.

Step 47. Create the common predictor tables

Make and Model are grouping metadata; they never enter these model features.

Step 48. Define numeric preprocessing

Median filling is a reusable safeguard; this verified CSV contains no missing predictors.

Step 49. Define categorical preprocessing

One-hot encoding creates indicator columns without imposing an artificial order.

Step 50. Create a numeric-model helper

This helper selects only the declared numeric inputs.

Step 51. Create a mixed-model helper

The helper creates a fresh fitted pipeline for each candidate and validation fold.

Step 52. Define fuel-consumption interactions

A declared fuel-code indicator multiplies consumption; it uses no target values.

Step 53. Create the constant baseline

The estimator ignores its input values and predicts the mean training target.

Step 54. Create the engine-only candidate

Use the same pipeline interface so all candidates share the same validation procedure.

Step 55. Create the engine-plus-cylinders candidate

This is the training-only version of the supplied notebook model.

Step 56. Create the specifications Ridge candidate

This candidate does not need a fuel-consumption rating.

Step 57. Create the fuel-only candidate

This tests whether one reported fuel rating already explains much of the target.

Step 58. Create the fuel-aware Ridge candidate

Ridge uses all declared numeric features and encoded fuel/class categories.

Step 59. Create the fuel-interaction candidate

Different fuel-specific consumption terms allow different slopes.

Step 60. Create the random-forest candidate

Settings are fixed demonstration choices, not a claim of optimal tuning.

Step 61. Create grouped validation folds

Validation groups also stay separate inside the training partition.

Step 62. Define the scoring metrics

Errors use g/km; R² compares squared error with predicting the validation mean.

Step 63. Run the validation comparison

Scikit-learn negates error scores because its scoring API treats larger values as better.

Step 64. Display the comparison table

Lower mean validation RMSE is the declared selection criterion.

Step 65. Select the fuel-known model

The winner is chosen without using any held-out test score.

Step 66. Select the specifications-only model

Select the strongest specifications-only candidate on the same training folds.

Step 67. Plot validation errors

The plot makes the improvement over simple alternatives easy to compare.

9. Measure performance on held-out vehicles

Fit the selected models on all training records, then make the reserved test predictions. We report the selected fuel-known model, the selected specifications-only model, the original two-feature approach and the mean baseline. The comparison is pre-declared; test errors do not choose the winner.

MAE is mean absolute error. RMSE emphasises larger errors. R² is a squared-error comparison with a mean prediction; 0.997 is not “99.7% classification accuracy”. A high score with fuel consumption available mostly reflects closely related physical ratings.

Step 68. Define one metric function

A shared function prevents accidentally scoring different targets or units.

Step 69. Fit the selected fuel-known model

Only training records enter .fit().

Step 70. Predict held-out ratings

These records have unseen exact Make–Model identities.

Step 71. Score the selected fuel-known predictions

Compare predictions with the reserved target in the original g/km unit.

Step 72. Fit the specifications-only winner

This model uses no fuel-consumption rating.

Step 73. Predict from specifications only

The fitted pipeline selects its declared inputs from the common test table.

Step 74. Score the specifications-only predictions

Use the same target and metric function as the fuel-known model.

Step 75. Fit the corrected two-feature comparator

This reproduces the v3 feature choice with a proper training-only fit.

Step 76. Predict with the two-feature comparator

This model uses only engine size and cylinders.

Step 77. Score the two-feature comparator

Its held-out error is distinct from the supplied v3 in-sample error.

Step 78. Fit the constant baseline

A score is more useful when a reader sees what a trivial prediction would achieve.

Step 79. Predict with the constant baseline

Every test record gets the mean of the training targets.

Step 80. Score the constant baseline

This gives the test-set reference for the more informative models.

Step 81. Compare held-out results

The table uses the same records, target and metric functions.

Step 82. Store residuals with the source identity

A positive residual means the model underpredicts the recorded rating.

Step 83. Plot actual versus predicted

The red diagonal is exactly predicted = actual, not a fitted best-fit line.

Step 84. Plot residuals versus predictions

Look for systematic trends, changing spread and exceptional observations.

Step 85. Summarise errors by fuel type

The sample count is part of the interpretation, especially for rare categories.

Step 86. Summarise errors by vehicle class

A class with very few held-out records provides weak evidence about future performance.

Check uncertainty and repeated-row sensitivity

These short extensions strengthen interpretation without hiding the basic workflow. The group bootstrap resamples held-out model groups and does not refit the selected estimator. Its interval is conditional on this fitted model and an exchangeable-group assumption; it does not include model-selection uncertainty or establish national representativeness.

The repeated-row sensitivity comparison uses training-only validation, leaving the held-out test results untouched. Keep the main run and the sensitivity run clearly labelled.

Step 87. Index the held-out groups for resampling

All records from a sampled model group move together.

Step 88. Run the group bootstrap

This estimates variability in RMSE across resampled held-out groups.

Step 89. Report the conditional interval

A percentile interval is different from the standard deviation between cross-validation folds.

Step 90. Create the deduplicated training view

Exclude tracking and audit columns when identifying repeated source records.

Step 91. Run validation on the sensitivity view

Use the selected candidate settings, not a new search driven by test performance.

Step 92. Compare the repeated-row policies

Removal changes observation weights and fold allocation, so this is sensitivity rather than a paired test.

Compare with combustion arithmetic

Natural Resources Canada reports approximately 2.3 kg CO₂ per litre of gasoline. That gives a rough gasoline-only rule:

The factor is approximate and not fitted to this CSV. We compare it with the selected estimator on the same X/Z gasoline rows. It does not apply to diesel, E85 or natural gas.

Source: NRCan personal vehicles.

Step 93. Select the comparable gasoline subset

Both methods must be scored on the same observations.

Step 94. Evaluate the gasoline benchmark

A simple scientific rule is a useful alternative to a learned model.

10. Business application: fleet operating scenarios

For a fleet manager, the useful question is often “What changes in annual fuel cost and tailpipe CO₂ if we change the vehicle mix?” The following demonstration uses catalogue class means as proxies, with full-row repeats removed for that summary. Those means are not matched replacement vehicles, current South African ratings, or actual fleet performance.

We compare 100 vehicles at 25,000 km per vehicle per year. The fuel price is an assumed R24/L, not a retrieved current price. All three fleet options have the same vehicle count and distance.

Step 95. Create a unique catalogue summary

This view is descriptive and is created after the completed model evaluation.

Step 96. Declare the fleet assumptions

Change one assumption at a time to see how the arithmetic responds.

Step 97. Calculate annual fuel use

The units are litres; dividing by 100 converts the consumption rating to litres per kilometre.

Step 98. Calculate the fuel budget

This is a scenario budget, not a quoted fleet operating cost.

Step 99. Calculate annual tailpipe CO₂

One tonne contains one million grams.

Step 100. Compare the two ends of the scenario

Differences are arithmetic consequences of these proxy ratings and assumptions.

Step 101. Plot the fleet scenario

Two charts keep currency and tonnes on separate axes.

How a business would use this

The inferred business value is a repeatable comparison: reconcile actual vehicles, add distance and fuel prices, then compare options under equivalent operating requirements. Purchasing decisions also require depreciation, maintenance, financing, charging/infrastructure and duty cycle. A vehicle supplier could use validated local ratings and forecast sales to compare product mixes and customer operating scenarios.

Under the GHG Protocol, combustion from owned/controlled vehicles can fall in Scope 1. A manufacturer's sold-product use phase is a different Scope 3 Category 11 boundary, normally assessed over expected lifetime use. The calculations here are tailpipe CO₂ only; they are not a complete CO₂e or lifecycle inventory and do not handle the special accounting of biogenic carbon in E85.

Sources: GHG Protocol Corporate Standard, Scope 3 calculation guidance, Category 11.

11. Business application: tax exposure is jurisdiction-specific

South Africa illustration. The SARS Schedule 1 Part 3D listing was checked on 4 October 2026 (listed update: 1 April 2024). For a qualifying passenger vehicle in this simplified example, use R146 per g/km above 95 g/km. The 2025 SARS policy confirms the certificate-based calculation. Actual liability requires the applicable classification, clearance date, recognised test report and any exclusions. We do not calculate VAT, other duties or non-passenger classifications.

The inputs below are assumed locally valid rated emissions. Historical Canadian catalogue ratings and machine-learning estimates are not substitutes for the required South African report. If the required test evidence is missing, the legal fallback applies; this model is not that fallback.

Sources: SARS current schedules, Part 3D, SARS motor-vehicle emissions policy.

Step 102. Declare the South African illustration parameters

Freeze the checked rate and threshold so the example can be reproduced.

Step 103. Calculate the taxable excess

Only the rating above the threshold enters this simplified levy formula.

Step 104. Calculate levy per unit and scenario total

This is base levy arithmetic, with all eligibility assumptions declared above.

Step 105. Measure sensitivity to a rating difference

A rating error can matter financially even when a model has a small average error.

Canada uses a different tax measure

Canada's Green Levy uses weighted fuel consumption: 55% city plus 45% highway. The official table is CAD 0 below 13 L/100 km, CAD 1,000 for 13–<14, CAD 2,000 for 14–<15, CAD 3,000 for 15–<16, and CAD 4,000 at 16 or more. Eligibility excludes pickups and certain other vehicles. It is not a universal per-gram CO₂ tax.

The illustration below uses assumed eligible passenger vehicles, not every row in this catalogue. Source: CRA current rates of excise taxes, checked 4 October 2026.

Step 106. Create the Canadian tax scenarios

The examples are assumptions selected to demonstrate the slab boundaries.

Step 107. Apply the Canadian rate bands

Do not round a rating into another band without the applicable measurement rule.

Keep the financial and environmental quantities separate

None of these amounts is atmospheric concentration in ppm. Do not add ZAR and CAD amounts, or confuse a one-vehicle fiscal charge with recurring tonnes of emissions.

12. Audit application: screen, trace, corroborate

An audit perspective asks whether records are complete, correctly measured, properly classified and supported by evidence. Model residuals can prioritise review, but cannot establish fraud, certify emissions, or independently verify the same physical rating system they were trained on.

We use training-only out-of-fold predictions to set a 99th-percentile screening threshold. The threshold is a declared analytics choice, not statutory materiality. It also reflects model selection on the training data and is not an independent confidence statement.

Step 108. Generate out-of-fold training predictions

Each training row is predicted by an estimator fitted without its model group.

Step 109. Set a training-derived review threshold

Do not choose a cutoff by inspecting the largest test errors.

Step 110. Create a held-out review queue

Keep source rows and vehicle identities so a reviewer can trace the input.

Step 111. Examine the largest exception

A numerical example shows why the review needs original supporting evidence.

Step 112. Review the arithmetic flags separately

A rating-formula discrepancy is a different control from a model residual.

What an auditor would request next

For the SA manufacturer example, SARS requires a report from an NRCS-recognised testing laboratory and retained evidence. A model estimate can support planning and exception screening; it cannot replace that evidence. Sources: SARS motor-vehicle policy, GHG Protocol Corporate Standard.

13. Country perspective: use sales weights and activity

To explore what vehicle types being sold could imply, begin with a sales/registration-weighted rating, not the mean of catalogue entries:

For a new-sales cohort driven a full assumed year:

N is vehicles, K is kilometres per vehicle and E is g/km. The following mixes are hypothetical. Class-average intensities come from the unique Canadian catalogue view; volumes and mileage do not. No actual national sales, emissions, or market response is estimated here.

Step 113. Declare two hypothetical sales mixes

Both mixes have 100,000 vehicles so a volume difference does not explain their comparison.

Step 114. Attach the proxy rating for each class

In a real analysis, join registrations to matching local year/variant ratings.

Step 115. Compute a sales-weighted rating

The weights are assumed vehicle counts, not catalogue record frequencies.

Step 116. Convert to full-year cohort tonnes

This is a full-year equivalent for the new cohort, not the observed year-of-sale total.

Step 117. Compare the hypothetical mixes

Hold distance and cohort size fixed while changing the assumed mix.

Step 118. Show the distance sensitivity

A lower rating does not guarantee a lower total if the fleet drives more kilometres.

Step 119. Plot the sales-cohort scenario

The title and axes explicitly identify the hypothetical nature and units.

What is required for a national estimate?

Statistics Canada publishes registration data by fuel/type/geography; matching classifications and years would be a next step, not an automatic join to this file. IPCC guidance estimates road CO₂ primarily from fuel consumption and carbon content and recommends reconciling activity information. Biogenic carbon needs separate treatment. A sales-cohort scenario supplements that work; it is not a replacement national inventory or a ppm model.

Sources: Statistics Canada registration table 20-10-0025-01, IPCC 2006 Guidelines, Volume 2 Chapter 3.

14. Summarise the value without overstating it

The strongest portfolio evidence is the combination of reproducible code and clear judgement: recognise the data's limits, choose validation that reduces repeated-model overlap, compare against simple alternatives, and connect metrics to a specific decision.

This notebook estimates recorded vehicle ratings. It does not certify new designs, prove fraud, calculate a named company's tax bill, establish premium-fuel causality, or measure a country's atmospheric CO₂ concentration. Current local certificates, activity and market data would be needed to use the workflow operationally.

Learning exercises

1. Explain why a model can have high R² but still be unsuitable for certifying a tax rating.
2. Change the fleet mileage assumption. Which outputs double if mileage doubles?
3. Change only the fuel-price assumption. Why do estimated tonnes stay the same?
4. Compare retained and removed repeated rows. What source information would settle the policy?
5. Identify why a low-emission catalogue mean is different from a sales-weighted mean.
6. Find the largest audit exception and list the original documents needed to resolve it.
7. Explain why Scope 1, lifetime sold-product use and a national territorial total cannot be added blindly.

For your website, lead with the question and verified results. Show the fuel relationship, validation comparison, residual plot and one labelled business scenario. Make full code available through this notebook or GitHub, and keep assumptions next to their scenario results.

15. Optional exports and run record

SAVE_OUTPUTS=False prevents the notebook from writing charts and tables. Set it to True before Run All to save the figures and the following traceable result files locally. Publishing the notebook does not automatically update a hosted website or tax-rate snapshot.

Step 120. Record the runtime versions

Versions, input hash and settings make a result easier to reproduce.

Step 121. Prepare a model-run record

Keep input provenance and selection information with the measured results.

Step 122. Export the principal result tables if enabled

These exports have clear names and do not alter the original CSV.

Step 123. Export the scenario tables if enabled

Each scenario retains its own units and assumptions.

References and provenance

Dataset supplied by the user: CO2 Emissions_Canada.csv. The v3 notebook cites Dovudkhon's Kaggle dataset, accessed 1 April 2024. The precise year range and licence are not independently verified here; model year is absent from the supplied CSV. Official NRCan data is background context, not a claim that this local file was downloaded directly from the government portal.

Technical sources

• scikit-learn: linear regression.
• scikit-learn: grouped cross-validation.
• scikit-learn: preprocessing and leakage.
• scikit-learn: regression metrics and baselines.
• NRCan: personal vehicles and gasoline CO₂.

Business, tax, audit and national context — checked 4 October 2026

• SARS: current Customs and Excise schedules, Schedule 1 Part 3D listing.
• SARS: Schedule 1 Part 3D.
• SARS: motor-vehicle emissions external policy, effective 30 June 2025.
• CRA: current rates of excise taxes, Green Levy bands and exclusions.
• GHG Protocol: Corporate Standard, reporting boundaries and verification context.
• GHG Protocol: Scope 3 calculation guidance, Category 11, expected use phase of sold products.
• IPCC: 2006 Guidelines, Volume 2 Chapter 3, road-transport methods and reconciliation.
• Statistics Canada: table 20-10-0025-01, new motor vehicle registrations.

The notebook proposes potential applications from those sources and the measured model results. It does not claim that the proposed workflow has been validated for statutory tax, external assurance or national emissions accounting.