A Garmin Connect IQ widget that predicts the weather using only your watch's barometer and compass - no phone, no internet required.
Forked from simonl-ciq/SimplyWeather. Original app used the Zambretti algorithm; this fork replaces it with the Sager Weathercaster engine, adds a regression-based barometric trend analyser (altitude-immune and diurnal-corrected), glance-view weather icons, and various quality-of-life improvements.
The forecast engine is based on Raymond Sager's meteorological method (1960s, US Navy). Unlike simpler barometric forecasters, Sager treats wind direction as a primary forecast dimension alongside pressure and its trend.
Inputs (all derived on-device):
- Current barometric pressure (hPa)
- Pressure trend over the configured time window (rising / steady / falling)
- Wind direction from the compass (8 octants)
- Current date (for continuous seasonal corrections)
- Hemisphere (north / south)
How it works:
- Three lookup tables (
steadyBase,risingBase,fallingBase) are indexed by wind octant (0-8), producing a base forecast number (0-25). - The base number is adjusted by pressure level (±2) - high pressure biases toward fair, low toward unsettled.
- A seasonal modifier (±1) accounts for summer convective storms and winter clearing patterns, ramping continuously with the date instead of stepping at month boundaries.
- The final forecast number maps to a condition label (e.g. "Fairly fine, showers likely") and a precipitation probability (0-95%).
26 forecast conditions range from Settled fine (0) to Stormy, much rain (25).
Provenance, honestly: the 26 condition labels are Zambretti's vocabulary, not Sager's, and the three lookup tables are hand-built rather than transcribed from Sager's published matrix. Full Sager takes five inputs - wind direction, wind-direction change, barometer reading, barometer change and present weather - and this engine feeds it two. The precipitation percentages have no published source. Neither method's original calibration survives that, which is what the measured skill below reflects.
The rising / steady / falling input to Sager is not a naive "now minus three hours ago" comparison - it comes from a small on-device regression pipeline run over the barometer's stored history:
- Mean-sea-level reduction. Every pressure sample is first reduced to mean sea level using the watch's barometric elevation history, gated so that only movement too fast to be weather counts as altitude: a climb, a drive uphill or an elevator cancels out and only genuine weather moves the trend. The pressure shown and the rain model use the same reduction, so the three cannot disagree. With no elevation history the trend falls back to raw station pressure, and the forecaster is told the reading is not a sea-level one.
- Quadratic regression. A single-pass least-squares parabola is fitted to the sea-level series across the trend window (~6 h). The fitted curve gives the net pressure change over the window, tested against a deadband (default 0.17 hPa/h, Sager's "slowly" boundary of 1 hPa in 6 h) to decide rising, steady, or falling. Fitting a parabola rather than a straight line lets a curving pressure profile be read correctly.
- Diurnal tide correction. Pressure rises and falls every day with no weather behind it - about 1-2 hPa at mid-latitudes, more in valleys and basins. The watch learns its own site's daily cycle, in 24 sun-time slots, from the pressure record the barometer keeps whether or not an app is running, so it keeps learning through activities, nights and days the glance is never opened, and subtracts it from every window. A latitude climatology stands in only until the record covers a day. Travel more than 200 km and the learned cycle is kept - most of it is solar, and in a traveller simulation across ten sites keeping it halved the leftover daily cycle on the day of arrival against starting over - while the windows that straddle the journey are skipped.
- Short-window front detection. An independent linear fit over only the last 3 hours flags a fast-moving front before the longer window catches it. Because it is computed from the recent samples alone, a large pressure wiggle earlier in the window cannot contaminate it.
- Hysteresis & front passage. The trend is quick to raise an alarm and slow to clear it, and a passing front (was falling, now levelling with pressure recovering) is upgraded to rising.
A glitched elevation sample cannot poison the series: the sea-level reduction clamps altitude to a physical range. This pipeline replaces an earlier point-sample second-derivative ("acceleration") trigger that over-reacted to short pressure wiggles and to altitude changes.
Earlier revisions of this file quoted accuracy figures of 65-85%. Those were never measured. They have been replaced with a verification against a rain gauge.
The engine was replaced in September 2026. The Sager lookup table it used before was scored over 73 days against a co-located weather station and found to separate wet hours from dry ones no better than a coin flip: area under the ROC curve 0.52, in every cross-validation fold. Its stated probabilities were not probabilities either - "75%" verified at 13%.
What ships now is a single calibrated feature: tide-free sea-level pressure measured against the site's own recent history, in units of the site's own recent spread. Scored over 73 days, 1400 day-ahead forecasts, 21.4% of 24-hour windows wet:
| Metric | New engine | Old table | Reading |
|---|---|---|---|
| Brier score | 0.158 | 0.226 | climatology scores 0.168 |
| Discrimination (fold AUC) | 0.73 | 0.52 | 0.5 is a coin flip |
| Reliability | 24% stated -> 29% observed | 75% -> 13% | the number now means something |
| Resolution | 0.011 | 0.001 | information about when |
| False-alarm ratio | 0.68 | 0.88 | still high; see below |
What it does and does not do. It orders hours well - given two hours, it reliably ranks the wetter one higher, and it does so in every fold of the record. It does not beat a constant forecast of the local average: out of sample its Brier skill score is +0.02, and a control feature consisting of nothing but "days since the record started" scores +0.03 on the same data. Seventy-three summer days cannot establish what the local average is across a year, so the absolute level is not yet earned. That is why the probabilities are narrow (5% to 60%) and why the descriptions above "Rain at times" are unreachable: a single barometer does not know enough to say Stormy.
Away from home. The same feature was scored at ten sites on four continents (NOAA ISD 2019-2023 plus the home station), each held out from the fit. The ordering travels to Europe - area under the curve 0.63-0.66 in the Alps, on the Atlantic and on the Black Sea - weakly to the American plains (0.54), and not at all to Sydney or the tropics (0.50-0.52). The level does not travel: the share of wet 24-hour windows runs from 8% in Phoenix to 63% in Brest, and a fit pooled over the other sites did worse on average and worse at home than the home fit. So the percentage is calibrated at home, and on the road it is a ranking: higher means wetter than usual for where you are, not a stated chance.
Removing the daily tide from the reading matters because the day means it is compared against have none: left in, the rain chance on the home archive carried a daily swing of its own, 21% at 09:00 sun time against 24% at 16:00; removed, the swing is 0.4 points.
The horizon moved from 6 hours to 24 for the same reason - the barometric signal at this site is a day-ahead one (rank correlation -0.68 at 24 h against +0.00 at 6 h).
The wind-bearing path is not separately verified: the reference station logs wind direction without long-term statistics, so there is no history to score it against yet. The old table is used only until 14 days have closed; the first pass closes every day the watch's own record already holds.
Read the forecast as a well-ordered barometer readout, not as a calibrated probability of rain.
- Two-line forecast text (e.g. "Fairly fine" / "possible showers early")
- Precipitation probability percentage with a rain / snow icon (season-aware)
- Pressure trend indicator: Rising, Steady, or Falling
- Current barometric pressure (hPa)
- Live compass display with 16-point cardinal directions (N, NNE, NE, ...)
- Heading smoothing and direction hysteresis to avoid jittery updates
- Wind direction is persisted across widget sessions
- Shake-to-recalibrate: shake the watch to reset the compass heading
A compact glance view with:
- Customisable title (configurable in Garmin Connect settings)
- Current forecast summary text
- Weather icon - context-aware by time of day and season (see table below)
The glance view selects an icon based on three inputs: the Sager forecast number, time of day, and season.
Day / night is determined by a fixed 07:00-19:00 window.
Season is hemisphere-aware - Northern: Dec-Feb = cold season; Southern: May-Sep = cold season.
| Forecast | Condition | Warm season (day / night) | Cold season (day / night) |
|---|---|---|---|
| 0-1 | Clear / fine | ☀️ Sun / 🌙 Moon | ☀️ Sun / 🌙 Moon |
| 2-6 | Fair / variable | 🌤 Cloud-day / ☁️🌙 Cloud-night | 🌤 Cloud-day / ☁️🌙 Cloud-night |
| 7-14 | Showers / unsettled | 🌧 Rain-day / 🌧🌙 Rain-night | 🌨 Snow-day / 🌨🌙 Snow-night |
| 15-21 | Rain / very unsettled | 🌧 Heavy rain | 🌨 Heavy snow |
| 22-25 | Stormy | ⛈ Thunderstorm | 🌨❄️ Snowstorm |
The main widget view shows a small raindrop (warm season) or snowflake (cold season) icon next to the precipitation percentage.
- Automatic hemisphere detection via GPS (one-shot fix)
- Falls back to the configured default (Northern or Southern) when GPS is unavailable
- Seasonal adjustments for precipitation type (rain vs. snow) and forecast modifiers
- Fenix 6 / 6 Pro / 6S / 6S Pro / 6X Pro
- Fenix 7 / 7 Pro / 7S / 7S Pro / 7X / 7X Pro
- Fenix 8 (43 mm / 47 mm) / Fenix 8 Solar (47 mm / 51 mm)
- Fenix Chronos / Fenix E
- Forerunner 965
Requires Connect IQ SDK 2.4.0 or later. Additional devices can be added via
manifest.xml.
| Permission | Reason |
|---|---|
| Sensor | Access barometer and magnetometer for pressure readings and compass heading |
| SensorHistory | Read barometric pressure history to calculate pressure trends |
| Positioning | Detect hemisphere (north/south) via GPS for seasonal corrections |
Configurable from the Garmin Connect app:
| Setting | Description | Default |
|---|---|---|
| Device pressure correction | Offset added to the barometer reading (hPa) | 0 |
| Trend threshold | Rate of pressure change below which the trend reads "steady" (hPa/h) | 0.17 |
| Trend time window | Hours of pressure history used to determine the trend | 6 |
| Display details | Show temperature and extended info on the widget face | Yes |
| Default hemisphere | Hemisphere fallback when GPS is not available | Northern |
| Glance title | Custom title displayed in the glance view | Weather |
- English
- German (Deutsch)
Build with the Garmin Connect IQ SDK and side-load the .prg file to your watch.
- Clone or download this repository.
- Open the project in Visual Studio Code with the Monkey C extension.
- Build for your device (
Monkey C: Build for Device). - Copy the generated
.prgfile to your watch'sGARMIN/APPSdirectory.
- Connect IQ SDK 2.4.0+
- Visual Studio Code with the Monkey C extension
# Build via the VS Code command palette:
# Monkey C: Build for Device
# or use the Connect IQ CLI:
monkeyc -f monkey.jungle -o SimplyWeather.prg -d fenix7source/
SimplyWeatherApp.mc # Application entry point
SimplyWeatherDelegate.mc # Input handling & compass interaction
SimplyWeatherForecast.mc # Sager Weathercaster forecast engine
SimplyWeatherView.mc # Widget layout, rendering & pressure logic
resources/
drawables/ # SVG icons (weather, compass, etc.)
strings/ # App name
forecast-strings/ # Forecast condition descriptions (26 outcomes)
point-strings/ # Compass point labels (N, NE, E, ...)
settings/ # Garmin Connect configurable properties
resources-deu/ # German localisation
resources-eng/ # English localisation
- Original app: Simon (simonl-ciq) - the foundation this fork builds on
- Sager Weathercaster: Based on Raymond Sager's barometric forecasting method (1960s, US Navy)
- Icon design: Freepik from Flaticon, licensed under CC BY 3.0
This project is licensed under the MIT License - see the LICENSE file for details.