BEV Trajectories · @LeRaffl

This repository powers the public BEV Trajectory Gallery, an interactive dashboard that visualizes electric-vehicle transition dynamics for countries around the world.

▶ Open the live gallery:
https://leraffl.github.io/LeRaffl-Gallery/#gallery

The gallery is updated continuously and provides, for each country:

• 📈 BEV transition trajectories (fitted logistic / Weibull-style curves)
• 🔄 ICE↔BEV market-share transition charts
• 🟫 TTM (trailing-twelve-month) market-split graphs
• ⏱️ Transition-time curves (speed of change between thresholds)

In addition, the dashboard includes two analytical views:

• Thresholds — When does a market reach 20%, 50%, 80% BEV share?
• Durations — How long does it take to move from one share level to another (e.g., 20→80%)?

All computations run purely client-side in the browser.

---

🚀 Project Purpose

Many countries follow an S-curve when transitioning from internal-combustion engines to battery-electric vehicles.
This project aims to make these transitions comparable, transparent, and publicly accessible, using a consistent modelling approach across all markets.

Inputs include:

• Monthly new-registration data (sources include ACEA, KBA, CPCA, Statistik Austria, and others)
• A generalized logistic (Weibull-like) model
• Harmonized modelling assumptions for cross-country comparability

Outputs:

• PNG charts for each country and month
• A parameter table (params.csv) containing all fitted model parameters
• Uncertainty bands per series (bands/<slug>.json: confidence, prediction and tolerance intervals)
• A machine-generated gallery manifest (manifest.json)

---

📂 Repository Structure

Key files in the repository root:

File	Purpose
index.html	The full interactive UI, single file, no build step. Four primary nav entries — Charts / Map / Rankings / Tools — with a sub-nav, over 14 sections (Gallery, World Map, Thresholds, Durations, Time Interval, Builder, Compare, Raw Data, Fleet, Data freshness, Submit Data, About, FAQ, Feedback). Runs 100% in the browser.
manifest.json	Auto-generated list of all available charts, used by the Gallery to display images.
build_manifest.R	Scans images/ and generates manifest.json.
params.csv	Contains the model parameters (v1, v2, t0, baseline year, last data month) for each market. Used for Thresholds & Durations.
bands/	Generated per render by R/bands.R: one <slug>.json per series with the 95 % confidence, prediction and tolerance bands around the fitted curve, plus the 10/20/50/80/90 % crossing years with their confidence ranges. Frontend data only (the PNGs do not use it): drawn as the optional uncertainty band in Builder and Compare. What the bands mean and how to quote them: docs/architecture/44-uncertainty-bands.md.
footnotes.csv	Optional curated per-country/variant chart footnotes (country,variant,footnote). render_country.R appends the matching note as a second caption line on the PNGs (e.g. Canada Whole's pre-2017 passenger-cars-only scope). Separate from the CSVs' internal notes column.
log_params.R	Produces and updates the parameter table (params.csv).
images/	Contains all exported PNG files structured as images/YYYY-MM/.... GitHub Pages serves them directly.
.github/workflows/build-manifest.yml	CI workflow ensuring that manifest.json always stays up to date.
docs/architecture/	Architecture handbook — components, data objects, interfaces, flows, secrets, ops. Kept up to date with every PR that changes structure. Start at docs/architecture/README.md.

The other data fetchers

The rows above are the ones that grew a long-form entry; they are not the full set. Data is fetched automatically for 30 countries in total — the ones without a row here are Albania, Austria, China, Colombia, Denmark, Finland, Ireland, Italy, Luxembourg, Malaysia, Netherlands, Poland, Portugal, Singapore, Spain and Sweden, plus the ~13-country ACEA cluster (scripts/fetch_acea.py).

A related but separately-curated set of countries gets Vans / HDV / Buses variants from a second, separate fetcher — scripts/fetch_acea_cv.py, ACEA's Commercial Vehicle press release — since ACEA publishes cars and commercial vehicles as two different reports, and which countries need which report differs (Germany/France/Sweden have their own passenger-car source but no national commercial-vehicle one, so they're in this roster and not the fetch_acea.py one; Switzerland left it in 2026-10 for ASTRA's monthly register counts; Greece stays in it although its passenger cars come from SEAA, because SEAA has no fuel split for commercial vehicles; see docs/architecture/38-source-acea-cv.md § 1a). ACEA only publishes cumulative year-to-date checkpoints (Q1/H1/Q1-Q3/full-year) for commercial vehicles, so this fetcher reconstructs genuine Q1/Q2/Q3/Q4 rows from them, falling back to a single yearly row only when no quarterly baseline exists yet for that country/variant/year; see docs/architecture/38-source-acea-cv.md for how.

For the complete, always-current picture — every fetcher, its source, the variants it writes and its schedule — see docs/architecture/02-components.md § 2.7 and the cron table in docs/architecture/08-deploy-ops.md § 8.11. Each country also has a public source page under sources/ explaining where its numbers come from.

---

🖼️ How the Gallery Works

When opening index.html, the browser loads:

1. manifest.json — the full list of charts
2. params.csv — all relevant model parameters
3. series/index.json, then series/<country>.json on demand — the observed monthly series behind Raw Data and the landing hero chart
4. Renders everything dynamically

The landing section opens with a hand-rolled inline-SVG chart for your home market, picked from your browser's timezone — no IP lookup, no location prompt — and a search box that writes straight through to the gallery filter (press Enter to jump to the gallery).

Gallery

• Search box on top (country, aliases, type, date, filename), country picker, chart types; always newest first
• Tag chips — Fit (Reliable / Unreliable / No transition), Status, Speed, Start — with live counts; selected filters show as removable chips above the results
• More filters: month and “Latest only”
• Show more adds about a screenful (24 charts) below the current ones without moving the view; Show all loads the rest
• Lightbox preview and direct download
• “Latest only” mode automatically selects the newest chart per type/country (there may be several even for a single month, e.g. when I'm cleaning up or trying something)

Thresholds

• Computes 20%, 50%, 80% share dates + any custom threshold
• Based on the fitted model parameters
• Exportable as CSV

Durations

• Estimates transition durations (20→80%, 10→90%, custom X→Y%)
• Includes the numerical speed at the inflection point
• Exportable as CSV

---

🔧 Data Generation (Overview)

Charts are generated locally using R scripts.
The workflow:

1. Monthly registration data is ingested and modelled
2. PNG charts are exported to the images folders by my locally run R scripts
3. build_manifest.R scans the directory → produces manifest.json
4. params.csv is carrying relevant output parameters from each model
5. GitHub Galerie shows updated gallery using the manifest as a guide of what to show with what filter
6. Github Thresholds and Durations are calculated on your device using the parameters from the R output cached in params.csv

The tables calculate thresholds and durations live.

Why isn't country X on the gallery? Some countries were investigated and deliberately left off because their freely-available data is incomplete (e.g. the dominant EV brand doesn't report), paywalled, login-walled, or because the market doesn't fit the new-registration / organic-transition model (e.g. Ethiopia's 100 %-EV import mandate over an ~85 %-used-import market). The rationale per country (Mexico, Ethiopia, … — plus Argentina and Colombia, which were shelved there first and have since been added from better sources) is documented in docs/architecture/14-data-source-gaps.md.

---

🧠 The model

1. What is being modelled

For each country, I model the BEV share of new car registrations over time. Empirically, this share does not behave like random noise. It behaves like a structured transition process: slow start, acceleration, eventual stabilization. The model is therefore not trying to “predict the future”, but to describe the structure of the transition as observed so far.

2. Why an S-curve at all?

Most large-scale technology transitions follow a similar qualitative pattern:

• early friction (costs, infrastructure, trust)
• positive feedback loops once adoption takes off
• diminishing returns as edge cases remain

This produces curves that are monotonic, bounded, and nonlinear. An S-shape is not assumed because it is pretty, it is assumed because it is the simplest structure that repeatedly matches real transitions. We've seen transitions between technologies in the past and this shape just fits.

3. Why not a normal distribution?

A normal distribution can, in fact, be fitted to cumulative adoption data like this. The problem is not that it can't be used. The problem is structure.

A normal distribution is:

• symmetric by construction
• unable to express asymmetric acceleration/deceleration
• unable to “break” as visibly when the data stops behaving like a transition

In other words: A normal distribution is just less suited.

4. Why a Weibull / generalized logistic formulation? The Weibull-style formulation is chosen for three reasons:

(1) Asymmetry Real-world adoption curves are almost never symmetric. The Weibull allows early-heavy, late-heavy, or roughly symmetric transitions — all with the same functional form.

(2) Failure visibility A Weibull does not HAVE to produce an S-curve. It could yield other shapes as well. It doesn't produce other shapes in most cases though, simply because the data fits S-shapes best. If the data does not support a meaningful transition, the model can:

• fail to converge
• produce nonsensical parameters (NaN, ±∞)
• collapse into degenerate shapes This is a feature, not a bug.

Example: Markets like Japan, which show no clear BEV transition, cause the model to break — exactly where it should. These examples are rare though. At time of writing I know of exactly 2 cases: Japan and Croatia

(3) Parametes The model uses two parameters. fewer → insufficient flexibility more → unstable estimation and overfitting Two parameters are a sweet spot, flexible enough to reflect reality, constrained enough to remain interpretable.

(4) Weights and calculation The basis for the fit was a basic Weibull distribution function of the form 1-exp(v[1]*x^v[2]). This function is being fitted to the data via a numerical OLS approach that introduces the market sizes as weights. These weights have turned out to be necessary in cases where countries have highly fluctuating market sizes. That is because BEV have proven to be much more stable in their absolute number of registrations than other fuel types, meaning that if in a certain month, there are much fewer newly registered cars, then this affects BEV much less than Petrol, Diesel or even Hybrids. As an aftereffect this means that with a relatively constant absolute number compared to this now shrunken total number, BEV share would produce an outlier that pulls the BEV share upwards more than it should. Weights correct this and make cases like Ireland much more stable. Convergence criteria are set to maxit = 100000 and reltol=10^-30.

5. What the parameters mean (conceptually) v1 — transition intensity Controls how aggressively the market moves once adoption starts. Intuitively: How sharp is the flip from “niche” to “mainstream”?

v2 — transition shape Controls where acceleration happens. Intuitively: Does the market ramp up early, or only after a long hesitation phase?

t0 — time shift It's a horizontal shift so the curve aligns nicely with calendar time. Ignore this one.

6. Most important model assumption Two hard assumptions are imposed:

• The transition starts at 0% (or 100% for ICE)
• The transition asymptotically ends at 100% (or 0% for ICE)

Technically these are assumptions. However, would we not assume this we'd introduce either less stable models or subjective models (which goes against what I want this to be). However, both the 0% and the 100% seem to fit reality quite well. Would they not fit reality, models would eventually output nonsense. If real data contradicts these bounds, the model stops fitting and this breakdown is visible. So far, mature markets (e.g. Norway or Denmark) genuinely converge towards 100%, not 70% or 80%.

7. ICE and PHEV curves Since the BEV curves fits so well, I have decided to redo an analogue curve for newly registered ICE cars as well. I struggled with the definition of what to call an ICE in this definition, but ended up at the split BEV & PHEV & ICE, meaning ICE in this categorisation includes HEV, H2, GAS, PETROL, DIESEL, ETHANOL and all other fuel types that are not BEV or PHEV (this also means that EREV are treated as PHEV). I could have chosen to define a Hybrid category for PHEV+HEV and sometimes I do even use it due to lack of more granular data (Türkiye), but this BEV-PHEV-ICE split seems like the one that yields the best insight on these trajectory curves. As a result of producing ICE and BEV curves, the leftover PHEV part should now not only make up the remaining percentage points to get to 100% if summed up, but if models behave, should also fit against real world data, so can be used as a sort of verification. Thus the graph showing 3 curves is produced.

8. This is a statement about TODAY This deserves to be absolutely unambiguous:

This is not a forecast. This is not a prediction. This is not a statement about the future.

The curve represents a best-fit description of the transition as of today. If future data changes the trajectory also changes, which is why I update it regularly. The parameters change, the curve changes, the derived thresholds change.

Think of it as: “Given everything we know right now, what would the transition look like if things simply continued?”

Not: “What will happen?”

9. Early-stage instability Uncertainty is not only a function of data quantity. It is also a function of where a market sits on the curve. Early-stage markets (low BEV share) are inherently volatile. Small absolute changes produce large parameter swings. This stabilizes naturally as the transition progresses. This stabilization is visible in the transition-time curves. Stability is also greater in bigger markets.

---

🌍 License & Usage

Charts and model outputs may be shared freely with attribution to @LeRaffl. Original data sources are subject to their respective licenses. Feel free to use these outputs for your own purposes, but please link me to whatever you do.

---