KODON WEATHER

01 / The gap

The national forecast doesn't know your valley.

Professional weather models (ICON, ECMWF, GFS) are extraordinary - at the scale of a 6-13 km grid box. But the decision you make happens at one point: this slope, this vineyard, this array. Between the grid and the point sits real money, and the generic forecast can't see it.

The generic forecast

×One number for a whole 6-13 km box
×Blind to your microclimate and terrain
×Same output shipped to everyone
×No idea what a miss costs you

✓Forecasts your exact station's readings
✓Learned on your site's own 9-year history
✓Tuned to the one decision that pays
✓Turns weather into euros saved

That gap is fine when the weather is just conversation. It's expensive when it drives a snow cannon, a spray rig, a grit truck, an energy bid, or a solar-market position. Those are point decisions, and they want a point forecast.

The scoreboard

What we beat - and who it's for.

Every row is measured on fully held-out data, against the real baseline it competes with. No strawmen, no simulation - and next to each, the customer who pays for it.

We beat	By	Who it's for
Operational ICON / GFS / ECMWF the global pro models, on 24 h rain probability	5 / 5 sweep	Events, construction, agriculture, logistics - anyone whose day is made or broken by rain.
The Davis Zambretti dial the 1915 algorithm every weather station ships with	22 / 25 wins	Every station owner running a 100-year-old forecast today.
Smart-persistence the field-standard solar baseline	+37% skill	PV operators & energy traders - tighter bids, fewer imbalance penalties.
Naive forecasting on temperature hourly, a full day ahead, MAE 2 °C	0.6 °C @ +1h	District heating plants & factory halls - schedule the technology around the real curve.
Flying blind on wet-bulb snowmaking window detection	802 h / yr	Ski resorts - stop paying to make snow that melts (€2-5K/night).
Missing a frost hard-frost detection, 6 h ahead	100% hit rate	Vineyards & orchards - one saved night is a rescued harvest.

The through-line: on the hyperlocal, decision-shaped variables, a €110 sensor + our engine out-predicts satellites and supercomputers - because they forecast a region, and we forecast your exact spot.

02 / The idea

A weather station is a decade of local truth. We forecast from it.

Every automatic weather station quietly records its own microclimate for years. That history is the training set no global model has: how this spot actually behaves. We learn one model per forecast step from it, benchmark honestly against the professional baselines, and only ship what beats them where it matters.

The substrate

One weather station + free open feeds (ERA5, NWP, satellite radiation), fused on a Raspberry Pi. Zero data cost, runs air-gapped on hardware the customer already owns.

The method (KODON / Helix)

Learned per-step models plus the KODON deterministic engine, cross-checked against smart-persistence and the operational NWP models. Honest-or-abstain: we report skill against the fair baseline, never a strawman.

The result isn't "better weather" in the abstract. It's a tighter number on the one variable a specific customer is paying to get right - and a measured claim about how much tighter.

03 / What it forecasts

Six decisions, six forecasts, one station.

Each product maps to a decision with a hard cost of being wrong. Every one is measured on held-out data (see the proof table below).

Energy · PV

Solar radiation

Global horizontal irradiance hours ahead - for PV output, market bids, grid balancing and battery control. +32% skill, sensor-only.

Ski resorts

Snowmaking windows

Wet-bulb temperature says exactly when water will freeze into snow that stays. Stop paying to make snow that melts.

Agriculture

Frost warning

Hard-frost detection at 100% hit rate on the test set. One saved night in a vineyard can be a whole season.

Municipalities

Road gritting

When to salt and when not to - F1 84%. Cut wasted call-outs without missing an icing event.

Utilities

Energy demand

72-hour min-temperature that drives heating and gas demand, calibrated to under 1 °C mean error.

Everyone

Rain probability

24-hour rain likelihood - the one variable where our engine sweeps the operational NWP models 5 of 5.

District heating & industry

24h temperature profile

Hour-by-hour temperature a full day ahead - MAE 0.6°C at +1h, ~2°C day-ahead. For heating plants and factory halls to schedule around their own microclimate.

04 / Proof

This isn't a pitch. It's measured, out-of-sample.

Every number below comes from a fully held-out test set on 9 years of real station data, benchmarked against the fair baseline - not a strawman. The figures are honest, and where a professional model still wins, we say so.

Product	Measured result	Baseline it beats
Solar (GHI)	+32% skill	Smart-persistence, held-out 2026. 30% daytime nRMSE - competitive with international ML solar forecasters, sensor-only.
Snowmaking	802 h / yr	Wet-bulb windows (Stull 2011) over 10 winters. Each avoided bad-window night saves €2-5K in energy.
Road gritting	F1 0.84	POD 85%, false-alarm 15%, 5-window held-out. Fewer wasted call-outs, icing events still caught.
Energy demand	0.997 °C	72 h min-temperature mean error, calibrated NWP - the driver of heating/gas demand.
Frost	100% hard-frost	Hit rate on hard frost, 6 h ahead. The event a grower cannot afford to miss.
Rain probability	5 / 5 sweep	Our KODON engine beats operational ICON / GFS / ECMWF on 24 h rain probability, same held-out data.
Temperature 24h profile	MAE 2.0 °C	Hourly temperature a full day ahead (0.6 °C at +1h). For district heating & factory planning - absolute accuracy, not just skill.

On raw temperature at synoptic scale, a well-placed operational model (ICON) is still stronger - and we say so plainly. Our edge is the hyperlocal, decision-shaped variables above.

05 / Watch it work

A held-out week, independently re-verified from the raw data.

The solar forecaster, checked line-by-line against the station's own 5-minute records - training counts, held-out split and baseline all reproduced from scratch.

verify-solar · vp2-kerm 2017-2026

# independent check, straight from 927,136 raw station records training records 2017-2024 : 776,565 (matches paper exactly) held-out test year 2026 : present, fully separated smart-persistence baseline : reproduced from scratch, honest # head-to-head on the held-out week learned model (XGBoost) : RMSE 72.9 W/m2 persistence baseline : RMSE 122.4 W/m2 skill vs smart-persistence : +32% (competitive, sensor-only)

Nothing here is asserted - it's re-runnable on the same machine the data lives on. And here is that week, drawn straight from the held-out test set:

Predicted vs actual solar radiation over the held-out test week, at four lead times — Predicted (green) vs actual (black) solar radiation, held-out test week 2026. Top row +1 h tracks the real curve including cloud dips; lower rows hold the daily shape a full day out.

Forecast error vs lead time — Error stays flat (~50-80 W/m2) across all lead times - far below both baselines.

Predicted vs actual scatter, all leads pooled — Predicted vs actual, all leads pooled - clustered on the 1:1 line.

Bonus - the 24h temperature profile Juraj flagged for district heating: the KODON forecast (green) tracks the real station temperature (white) across the day, issued a full day ahead. MAE ~2°C.

06 / How it makes money

Near-zero cost to run, a hard number saved per customer.

The economics are unusually clean: the data is free and it runs on a Raspberry Pi, so cost-to-serve is almost nothing. Revenue comes from one measurable saving per product. Ordered by how easily the buyer can do the math themselves:

Wedge	Buyer	The saving they can price today
Snowmaking Start here	Ski resorts	€2-5K / night
Solar / PV	PV operators, energy traders	Tighter market bids, fewer imbalance penalties, better battery dispatch - per site, per day.
Frost	Vineyards, orchards	One saved night = a rescued harvest. Insurance-grade value per alert.
Road gritting	Municipalities, road authorities	Fewer wasted grit call-outs at F1 0.84, without missing an icing event.
Energy demand	Utilities, gas suppliers	Sub-1°C 72 h min-temp sharpens demand and procurement forecasting.

Why the unit economics work

Cost to serve
~€0
Free data + a Raspberry Pi. No cloud GPU, no satellite, no per-query model cost. Gross margin is structurally high.
Saving per customer
Hard
Snowmaking especially: the buyer multiplies avoided bad-window nights by their own energy bill. The pitch proves itself.
Go-to-market
1+1+1
Land one paying pilot per product - snowmaking first - as a per-station subscription, then scale station by station.

Honest on market size: earlier internal TAM figures (€90M Slovakia, €2-3B EU) are un-validated planning guesses and we treat them as such. What is real is the unit economics above: near-zero cost to serve and a saving the customer can compute themselves. That is what we sell on.

For scale reference, not parity: Tomorrow.io, a weather-intelligence company built on satellites and ~500 staff, is valued above $1B. KODON WEATHER attacks the same value - the decision, not the map - from one sensor and a Pi.

07 / How it works

Station in, decision out. Four layers.

Data falls through the stack from raw readings to a priced decision, all on one owned box.

The stationOne automatic weather station (e.g. Davis VP2) logging its own microclimate every 5 minutes - the local truth no global model has.Live

Free feedsERA5 reanalysis, operational NWP and satellite radiation, all keyless and free, fused as extra features.Live

KODON / Helix enginePer-step learned models + the deterministic KODON core, benchmarked honestly against every baseline. Exact-or-abstain.Live

Decision layerTurns the forecast into the one call that pays: make snow / don't, grit / don't, bid / hold, alert / stand down.Per product

The whole stack runs on a Raspberry Pi, air-gapped, next to the station. The customer's data never leaves their site, and every forecast is an auditable, re-runnable computation.

08 / Roadmap

From one station to a network.

Done

Six measured products on one station

Solar, snowmaking, frost, gritting, energy and rain - all validated out-of-sample on 9 years of Kerm-station data.

Now

First paying pilot

Snowmaking at one ski resort, where the saving is hardest to argue with. Per-station subscription, prove the euro.

Next / months

Multi-station fusion

3-5 stations at different elevations vote per snow-gun / per-sector, so a whole resort or valley is covered, not one point.

Horizon

The point-forecast network

One product line, many stations, many verticals - each site paying for the one decision it cares about, all on cheap owned hardware.

09 / Integrity

We publish the claims that failed, too.

Retracted by us, before anyone asked: an early "world-rank #1 temperature" claim was withdrawn - it came from a unit-mixing bug and test-set leakage on our side. The honest result: a well-placed operational model still beats us on raw temperature; our real edge is the hyperlocal, decision-shaped variables. The internal TAM numbers are marked un-validated on purpose.

We measure everything, including ourselves. Our own from-scratch 0-LLM engine cut solar error in half (RMSE 122 → 83) - and where a tuned off-the-shelf model still edges it, we ship that one. We optimise for the best number on the box, never for whose it is. That discipline is exactly why every figure here holds up.

And what we won't claim: long-range prediction (physically impossible), replacing the global forecasting systems, or beating them at synoptic scale. A product whose whole value is an honest number has to be honest about its own.

"We don't sell a better map. We sell the one decision the map can't make for you - and we prove the saving."

Koscak Research

KODON WEATHER

The national forecast doesn't know your valley.

The generic forecast

KODON WEATHER

One station. Your weather.

What we beat - and who it's for.

A weather station is a decade of local truth. We forecast from it.

Six decisions, six forecasts, one station.

This isn't a pitch. It's measured, out-of-sample.

A held-out week, independently re-verified from the raw data.

Near-zero cost to run, a hard number saved per customer.

Why the unit economics work

Station in, decision out. Four layers.

From one station to a network.

We publish the claims that failed, too.

Hyperlocal weather intelligence
you own.

The national forecast doesn't know your valley.

The generic forecast

KODON WEATHER

One station. Your weather.

What we beat - and who it's for.

A weather station is a decade of local truth. We forecast from it.

Six decisions, six forecasts, one station.

This isn't a pitch. It's measured, out-of-sample.

A held-out week, independently re-verified from the raw data.

Near-zero cost to run, a hard number saved per customer.

Why the unit economics work

Station in, decision out. Four layers.

From one station to a network.

We publish the claims that failed, too.

Hyperlocal weather intelligenceyou own.

Hyperlocal weather intelligence
you own.