Bergen vs Paris Weather Comparison
This project compares 10 years of daily weather data (2016–2025) between Bergen, Norway and Paris, France.
- Bergen is on average 5°C colder than Paris, but the gap shrinks to ~4°C in winter
- Daily temperatures are 81% correlated — when one city gets colder, so does the other
- Bergen's snow days average 1.5°C — almost eight in ten are above freezing
- Bergen is geographically closer to the French Riviera than to Arctic Norway
- When Paris bakes above 30°C (114 days over the decade), Bergen averages a comfortable 18°C
Explore the data
The process
This project follows the School of Data pipeline methodology: Define, Find, Get, Verify, Clean, Analyse, Present.
Define
Started with a personal intuition: "Bergen and Paris weather are similar." We broke this into four testable hypotheses:
- H1: Daily mean temperatures are within 3°C of each other
- H2: Temperature changes correlate — when one city cools, so does the other
- H3: Snow events co-occur more than chance would predict
- H4: Sunny days co-occur more than chance would predict
Find
Surveyed six data sources (Open-Meteo, Meteostat, Frost API, Météo-France, Visual Crossing, NOAA GSOD). Open-Meteo was the only one with all six required variables (temp mean/min/max, precipitation, snowfall, sunshine) for both cities at 100% completeness, with no API key needed.
Get
Fetched 3,653 days of daily data per city from the Open-Meteo Historical Weather API (ERA5/ECMWF IFS reanalysis). Zero gaps, zero nulls.
Verify
- Completeness: continuous date sequences, no missing values
- Quality: no temperatures outside plausible climate records, no negative precipitation, no suspicious flat-line periods
- Cross-check: compared ERA5 vs ECMWF IFS models. Paris data agrees within 1°C on 94% of days. Bergen shows ~1.6°C model uncertainty — noted but not disqualifying
Clean
Merged both cities into a single horizon table (13 columns, daily granularity). Standardised units (sunshine to hours, snowfall to mm). Initially flagged extreme days (Paris max >35°C, Bergen min <−10°C) but a sensitivity check showed removing them changed results by <0.5°C, so we dropped the filtering.
Analyse
The analysis tested each hypothesis using simple statistical methods: daily temperature differences and their distributions (H1), Pearson correlation with lag analysis (H2), and co-occurrence rates compared to chance baselines (H3, H4). No modelling or machine learning — the dataset is clean and the questions are straightforward enough for descriptive and inferential statistics.
| Claim | Finding | Conclusion | |
|---|---|---|---|
| H1 | Daily temps within 3°C | Average gap is 5.1°C; only 25% of days within threshold | Rejected |
| H2 | Temperatures correlate day-to-day | r=0.81 daily, r=0.94 monthly | Confirmed |
| H3 | Snow events co-occur | 2.6× more than chance | Weakly confirmed |
| H4 | Sunny days co-occur | 1.4× more than chance | Weakly confirmed |
The H1 rejection turned out to be the most interesting finding. The 5°C gap hid a seasonal story: it shrinks to just 4°C in winter (with some months where Bergen was actually warmer than Paris) and widens to 6°C in summer. This triggered an Analyse→Define loop — the original intuition of "similar temperatures" was probably specific to colder months, where the difference is least noticeable, and had been extrapolated to the whole year. The reframed question became: Bergen is not much colder than Paris when cold matters most (winter), and the gap in summer is actually a feature — Bergen avoids Paris's heatwaves entirely.
An additional threshold sensitivity lesson: Bergen "has snow 20% of days" using the raw model data (any snowfall >0mm), but that includes trace amounts from the reanalysis model. Using a 5mm+ threshold ("noticeable snow") gives 14% — which matched lived experience much better.
Present
The goal was to help dispel a common misconception — especially among French people — that Norway is a country defined by snow and extreme cold. The site is structured in two parts: the what and the why.
The what explores the temperature comparison through both objective and subjective angles. The objective angle looks at the raw data: monthly averages, daily correlations, year-by-year heatmaps. The subjective angle reframes the gap as a lived experience: Bergen's snow days are warmer than you'd think, and its summers sit in the thermal comfort zone while Paris overshoots into heatwave territory.
The why explores geography — the distances between Bergen, Paris, and Tromsø — to explain why Bergen's climate is so different from the Arctic Norway people imagine. The Gulf Stream and Norway's sheer length are the key factors.
Many charts include a small interactive component to keep them engaging: filtering the heatmap by similarity level, scrubbing through years on the daily chart, or choosing a European capital on the flight distance map to make the "Norway is a long country" point personal and relatable.
The site is bilingual (EN/FR), auto-detecting French users, because the French audience is the primary one for this misconception.
Tooling summary
The technical implementation was built with Claude Code, following a guardrail-first methodology: each task delegated to the LLM was bounded by a single pipeline phase, and its output was verified before moving to the next step. Every phase produced a verifiable artifact — a script, a dataset, a report — so the entire pipeline can be reproduced without any LLM involved, just by running the Python scripts in order. The methodological structure followed the School of Data pipeline, providing the framework that guided each step.
Data pipeline — five Python scripts, each handling one step:
fetch_data.py → data/raw/ # Download from Open-Meteo API verify_data.py # Sanity checks + cross-validation clean_data.py → data/clean/ # Merge and standardise analyse.py # Test hypotheses export_for_web.py → site/data/ # Export JSON for the website
Website — a single-page static site with no build step:
- Chart.js for interactive charts (lollipop, time series, scatter, bar)
- TopoJSON + HTML Canvas for the background map and flight distance map
- Natural Earth / world-atlas for country boundaries (50m resolution)
- Vanilla JavaScript for i18n, animated transitions, and computed statistics
Publication — hosted on GitHub Pages, served directly from the repository root. Source code at github.com/clombion/bergen-paris-weather.
Sources
Data source — Open-Meteo Historical Weather API (ERA5 / ECMWF IFS reanalysis), 3,653 days of daily data per city. Snow days defined as days with 5mm+ snowfall. Thermal comfort zone based on Cheung & Jim, 2019.
Project data — the raw API responses and cleaned horizon table are stored in the repository for reproducibility:
- data/raw/ — raw JSON from Open-Meteo (Bergen + Paris, 2016–2025)
- data/clean/horizon.csv — the merged horizon table (3,653 rows × 13 columns)
- site/data/ — pre-computed JSON subsets powering the website charts