Pollution as a Proxy:
How NO₂ Data Predicts Industrial Output

In 2020, as governments imposed lockdowns, economists scrambled for real-time indicators. Official GDP figures come with a 6-week lag. PMI surveys miss the first week of a shock. But satellite-derived NO₂ data from Copernicus CAMS dropped instantly — and tracked the economic collapse and recovery almost hour by hour.

Hedge funds and macro desks have quietly known this for years. NO₂ — nitrogen dioxide — is primarily emitted by combustion: trucks, factories, power plants, aircraft. Where there is economic activity, there is NO₂. Where there isn't, there isn't.

The COVID natural experiment

The 2020 lockdowns provided the cleanest economic experiment in modern history: an abrupt, enforced halt to transport and industry across dozens of countries simultaneously. CAMS reanalysis data shows the signal clearly.

The contrast across cities is striking. Madrid saw the sharpest drop (−42%) — Spain imposed one of the strictest total lockdowns in Europe, halting nearly all non-essential industry. Milan followed at −35%, consistent with Italy's early hard shutdown of the Po Valley industrial corridor. Paris fell −34%; London, which maintained more essential activity, dropped −25%. Warsaw is the outlier: NO₂ barely moved (+4%) during March–April 2020 compared to 2019. Poland's lockdown was lighter and later, and a cold spring meant continued coal-fired residential heating. This isn't noise — Warsaw's result directly confirms what NO₂ measures: actual combustion activity, not just mobility. Where restrictions didn't cut industrial output, NO₂ didn't fall.

The recovery signal

Equally valuable is the recovery: NO₂ rebounded as lockdowns lifted, with a distinct pattern between countries with V-shaped vs. L-shaped recoveries. The chart below shows weekly mean NO₂ for Paris through 2020, alongside the IHS Markit France PMI.

The correlation is striking: NO₂ leads PMI by approximately 1–2 weeks during both the collapse and the recovery. By the time the April PMI print came in at 31.5 — a historical low — the NO₂ data had already been showing the same depth of contraction for three weeks.

Ten-year trend: structural shift or cyclical noise?

Looking beyond 2020, the longer-term NO₂ trend across European cities tells a more complex story about the energy transition, deindustrialisation, and vehicle fleet composition.

Three structural observations emerge from a decade of data:

• Paris and London show a consistent downward trend — Paris at −0.78 µg/m³/year, London at −0.49 µg/m³/year (OLS via aggregate="trend") — driven by the Euro 6 vehicle transition, natural gas replacing coal in power, and long-run deindustrialisation of city cores.
• Warsaw shows a fundamentally different pattern — NO₂ rose from 13.4 µg/m³ in 2013 to a peak of 20.4 µg/m³ in 2015, before partially recovering to 13.6 in 2024. The net OLS trend is nearly flat (+0.08 µg/m³/year), unlike the structural decline in Western Europe. Coal-fired residential heating and slower EV adoption set Central European cities apart.
• 2020 is a permanent reset, not just a dip. Post-COVID levels in Paris and London did not recover to 2019 peaks — remote work, reduced commuting, and accelerated EV adoption locked in a lower baseline.

The investment signal

How does this translate into actionable intelligence for investors?

Quantifying the edge

The value of alternative data is usually measured in days of lead time. For NO₂:

A 4-day lag versus a 23-day lag is a significant edge in fast-moving markets. For macro funds trading European rate volatility around data releases, knowing the industrial pulse three weeks early is not just interesting — it's potentially alpha-generating.

Getting the data

All the analysis above can be replicated with the Jiskta Python SDK in a few lines. For cities where an OSM administrative boundary is indexed, the area= parameter is the cleanest approach — no manual bbox needed:

from jiskta import JisktaClient
import pandas as pd

client = JisktaClient("your_key")

# Monthly series via OSM administrative boundary
df = client.query(
    area="paris",             # resolves OSM boundary automatically
    start="2013-01", end="2024-12",
    variables=["no2"],
    aggregate="area_monthly", # one row per month, area-averaged
)
# df: year_month, no2_mean — 144 rows, returned in ~180ms

# For London (use bbox until Greater London is in the OSM index)
df_london = client.query(
    lat=(51.3, 51.7), lon=(-0.5, 0.3),
    start="2013-01", end="2024-12",
    variables=["no2"],
    aggregate="area_monthly",
)

# OLS trend — one call, zero manual calculation
trend = client.query(
    area="paris",
    start="2013-01", end="2024-12",
    variables=["no2"],
    aggregate="trend",
)
slope = trend["slope"].mean()
r2    = trend["r2"].mean()
print(f"Paris NO₂ trend: {slope:+.3f} µg/m³/yr  (R²={r2:.2f})")
# → Paris NO₂ trend: -0.779 µg/m³/yr  (R²=0.09)

For multi-city portfolio coverage across an entire decade: geographic_tiles × months × variables = ~4 × 144 × 1 ≈ 576 credits per city — roughly €0.59 at Starter pricing. A portfolio covering 20 major European cities costs under €12.

Limitations and caveats

NO₂ is not GDP. A few important caveats for any financial application:

• Weather matters. Cold temperatures and atmospheric inversion trap NO₂ near the surface, creating spikes unrelated to economic activity. Always join ERA5 boundary layer height to adjust for meteorology.
• CAMS interim vs. validated data. The 4-day lag applies to interim reanalysis data, which is slightly less accurate than validated reanalysis (available with an 18-month lag). For real-time monitoring, use interim; for long-term trend analysis, use validated.
• Spatial resolution. A 0.1° grid cell (~11 km) covers mixed land use. A manufacturing district and a residential park may be in the same cell. Subsetting to known industrial zones improves signal quality significantly.
• This is not financial advice. All correlations shown are illustrative. Past correlation between NO₂ and economic indicators does not guarantee future predictive value.