Saratoga Wind Evaluator

Workflow Mode Selection

#1 Standard Workflow #1 Standard

#2 Advanced Workflow #2 Advanced

Select Forecast Location

Start Date

End Date

Model ?

Start Hour

End Hour

Races (1-6)

Model Evaluation

📊 Weather Model Validation & Consensus Engine • HRRR, NAM, GFS, ECMWF, GEM, ICON

🏆 Option 1: Robust Cached-Weight Consensus (Steps 7 & 8) Recommended (Top)

Blends forecasts from all 6 models using performance weights saved in your browser cache. Instantly loads using a quick parallel download.

Current Saved Model Weights ?

Equal Weights (Default 16.7% each)

⚖️ Priority Weight Overrides (Advanced) Show/Hide

Validation Period

Validation Baseline ?

Compare with Consensus ?

⚠️ Option 2: Real-Time Live Validation & Consensus (Alternative) Less Robust (Bottom)

Downloads 7 days of actual history and past forecasts, evaluates them in real-time to calculate weights, downloads current forecasts, and averages them in a single synchronous operation. Heavy on data, slower loading time (~6-8s), and subject to recent archive data availability.

📖 Venue Profile & Local Sailing Knowledge

🌬️ Race Day Forecast (Best Model)

📈 Wind Speed & Gusts (MPH)

🧭 Wind Direction (°)

🌬️ Race Day Forecast Comparison

📈 Wind Speed & Gusts (MPH) Selected Model

📈 Wind Speed & Gusts (MPH) Consensus Blend

🧭 Wind Direction (°) Selected Model

🧭 Wind Direction (°) Consensus Blend

🌡️ Conditions & Barometric Pressure Trends

📈 Temperature Trend

💧 Humidity & Precipitation Trend

📊 Barometric Pressure Trend

🌡️ Conditions & Pressure Comparison

📈 Temperature Trend Selected Model

📈 Temperature Trend Consensus Blend

💧 Humidity & Precipitation Selected Model

💧 Humidity & Precipitation Consensus Blend

📊 Barometric Pressure Selected Model

📊 Barometric Pressure Consensus Blend

🔥 NWS Fire Weather / Convective Data

🏔️ Atmospheric Stability Indicators (Mixing Height & Wind Profile)

⚡ Convective Analysis

Run analysis to populate mixing height inversion detection, gust potential, and wind shear calculations.

🧭 Wind Direction Trend & Model Agreement

↻ Veering / Backing Analysis

Run forecast to analyze direction trends and model agreement.

🧭 Direction Trend Comparison

↻ Selected Model Trend

Run forecast to analyze.

↻ Consensus Trend

Run forecast to analyze.

💨 Gust Direction Shift Analyst

↻ Veering vs. Backing Gust Detection

Run analysis to detect right-shifting (veering) vs. left-shifting (backing) gust patterns.

💨 Gust Direction Shift Analyst Comparison

↻ Selected Model Puffs

Run analysis to compute.

↻ Consensus Puffs

Run analysis to compute.

🌍 Map & Synoptic Overview (Morning to Evening) • HRRR (3km)

Run forecast to generate synoptic-scale analysis and planetary boundary layer wind shift projections.

🌍 Map & Synoptic Overview Comparison

🗺️ Race Area Map Selected Model

🗺️ Race Area Map Consensus Blend

📋 Synoptic Overview Selected Model

Run forecast to generate selected model overview.

📋 Synoptic Overview Consensus Blend

Run forecast to generate consensus overview.

🏁 Race Day Hourly Breakdown

Run analysis to generate race-day hourly breakdown.

🏁 Race Day Hourly Breakdown Comparison

Selected Model

Run analysis to generate breakdown.

Consensus Blend

Run analysis to generate breakdown.

⛵ Dynamic Race-by-Race Forecast Guides • HRRR (3km)

Select race window start/end times and count, then run forecast to generate individual race strategy guides.

⛵ Dynamic Race-by-Race Forecast Guides Comparison

Selected Model Guides

Run forecast to generate guides.

Consensus Guides

Run forecast to generate guides.

🧠 Geographic Strategy Engine

Run analysis to generate wind-direction-specific geographic strategy for the selected racing area.

🧠 Geographic Strategy Engine Comparison

Selected Model Strategy

Run forecast to generate strategy.

Consensus Strategy

Run forecast to generate strategy.

📊 After-Race Assessment (Optional)

📝 Log Today's Race Conditions & Performance

Evaluate forecast accuracy and log tactical observations. This form is optional and does not block forecast checks. Data is saved locally in your browser.

Which Forecast Elements Were Accurate?

Wind Speed (Predicted speeds matched actual conditions) Wind Direction (Mean directions & oscillations matched) Weather Forecast (General weather & frontal passage matched) Wave/Chop Forecast (Steepness & swell matched bathymetry) Local Current / Tide prediction matched water movement Centerboard & Rudder Weed-Clearance Diligence Practiced

More Accurate Weather Model

Favored Side of the Race Course

Skipper Notes & Tactical Observations

Logged Performance Summary

--%

HRRR Model Favor Rate

--%

GFS Model Favor Rate

Most Favored Side

--%

Weed Diligence Rate

Date / Venue	Best Model	Favored Side	Accuracy Checks	Weeds / Diligence	Skipper Notes
No after-race logs recorded yet.

Saratoga Wind Evaluator • Sailboat Racing Wind Forecast Evaluator • 2026

💡 Step-by-Step Wind Evaluator Guide ×

Follow these steps to analyze weather and strategy for your next regatta:

📋 #1 Standard Workflow Sequence

For quick forecasts using a single chosen model:

① Location Setup → ② Select Model → ③ Set Dates/Hours → ④ Get Forecast.

🔥 #2 Advanced Workflow Sequence

For consensus blending and side-by-side comparison screen splits:

①-④ Basic Setup (Select custom/preset, setup dates & hours). ⑤ Enable Model Evaluation (Toggles advanced validation panel). ⑥ Select Period & Baseline (Set validation range and baseline source). ⑦ Re-Calculate & Save Weights (Calculates model performance accuracy). ⑧ Get Robust Consensus Forecast (Averages forecasts based on weights). ⑨ Compare with Consensus (Splits maps & charts side-by-side).

⚠️ Sequence Restraint: You must Re-Calculate Weights (⑦) before running the Consensus Forecast (⑧) or enabling Comparison Mode (⑨).

① Forecast Location Setup
Select a preset venue from the dropdown, or select Other Location and search a city name. Custom coordinates, elevations, and depths will load automatically.

② Select Forecast Model
Choose which weather model to load. The default is the 4-Model Local High-Res Blend, which dynamically aggregates HRRR (35%), NBM (30%), HRDPS (25%), and ECMWF (10%) to output highly accurate localized racing forecasts.

③ Select Date & Hour Parameters
Specify start/end dates, active racing window hours, and the race count (used to compute Portsmouth handicap speeds).

④ Run Weather Forecasts
Click **Get Forecast**. The system queries NOAA NWS (with automatic fallback outside the USA) and Open-Meteo models, drawing wind and pressure graphs.

⑤-⑨ Advanced Consensus Steps
Under Advanced workflow, enable Model Evaluation (⑤), pick a range and comparison baseline (⑥), click Re-Calculate & Save Weights (⑦) to evaluate recent model runs, then get Robust Consensus (⑧) or toggle Compare with Consensus (⑨).

⑩ Log After-Race Assessments
After sailing, log which model (HRRR vs GFS) was more accurate and which side local sailors favored. Submitting updates local statistics and logs.

⑪ Cache Reset & Stale Data Recovery
If the forecast seems stale (showing identical plots after date/model updates) or offline states occur, click **Reset Cache**. This purges WebView storage and forces a clean re-download of all model runs.

📊 Model Validation Metrics Guide ×

This table compares forecast models against historical archive observations to check how accurate they have been recently. Here is what each metric means and how to evaluate them:

1. Speed MAE (Mean Absolute Error)
The average size of wind speed errors in MPH (regardless of whether the forecast was too high or too low). Lower is better. A MAE of 1.5 MPH means the model is, on average, 1.5 MPH off.

2. Speed RMSE (Root Mean Squared Error)
Similar to MAE, but it squares errors before averaging, penalizing larger errors heavily. A model with a low MAE but a high RMSE had a few massive "bust" forecasts. Lower is better.

3. Dir MAE (Direction Mean Absolute Error)
The average difference between forecasted and actual wind direction in degrees. Extremely important for sailboat racing to gauge how reliable a model is at predicting shifts. Lower is better.

4. Vector RMSE (Vector Root Mean Squared Error)
The primary standard metric. It combines speed and direction errors into a single mathematical 2D vector offset. This metric is used to determine the validation winner (highlighted row with 🏆). Lower is better.

5. Data Pts
The number of hours analyzed during the selected validation period. More data points make the validation metrics more reliable.

💡 How to Evaluate It All:

Winner (🏆): The model with the lowest Vector RMSE is marked as the winner. This model has been the most reliable over the validation period.
Consensus Weights: When you calculate consensus forecasts, models with lower Vector RMSE (better accuracy) are automatically given higher weights.
Custom Strategy: If a model has a very low Dir MAE, it is excellent for predicting wind shifts, even if its wind speed (Speed MAE) is slightly less accurate. Use this detail to pick your primary model!

♻️ Cache Reset Guide ×

The Reset Cache button performs a hard purge of the browser's storage and PWA cache.

What it does:
1. Deletes the Progressive Web App (PWA) cache storage containing code files.
2. Unregisters browser Service Workers.
3. Clears local and session storage (including cached weather model weights).
4. Forces a cache-bypassing page reload.

When to use it:
* Stale Data / Identical Plots: If the graphs continue to display the same or outdated curves after changing the date or model.
* Code Updates (Browser): After applying updates or fixes to the application code, to ensure your browser is running the newest files.
* Inconsistent State: If a network error causes the app's internal variables to get out of sync.

Mobile App / Android Behavior:
* Asset Packaging: Under Android, all HTML, CSS, and JS assets are bundled directly inside the APK package. This button clears your local weights database and WebView cache, but code updates are loaded from the package files. * Applying Updates: To receive future code changes on your phone, you must install/sideload the compiled debug APK. The app will then always run the latest code packaged in the APK.

💡 Weather Models Evaluation Guide ×

Summary of model strengths, weaknesses, and biases based on historical evaluations for Saratoga Lake:

⚠️ Forecast Horizon Reliability Rule: The farther out a forecast window is, the less reliable the predictions become. High-resolution local models (HRRR/HRDPS) focus strictly on near-term details (18–48 hours) to maintain microscale accuracy, while global models (GFS/ECMWF) extend much farther (10–16 days) but degrade in reliability.

Model	Resolution / Max Forecast Lead Time	Strong Points / Best Use	Weak Points / Biases
High-Res Blend	2.5–9 km / 48-hour range / 4-Model Weighted Blend	Combines 35% HRRR (hourly shifts/thermal development), 30% NBM (bias-corrected sustained speeds), 25% HRDPS (local terrain channeling), and 10% ECMWF (synoptic anchor to stabilize timing errors). Absolute best general default for Saratoga Lake racing.	Requires loading four separate models (slightly higher data usage). Range is limited to 48 hours.
HRRR	3 km / 18–36 hour range / Hourly updates	Excellent for near-term hourly wind shift trends, thermal mixing, and frontal transition timing.	Tends to overpredict light-air wind speeds (by 1-3 mph). Needs NBM correction.
HRDPS	2.5 km / 48-hour range / 6-hr updates	Superior terrain resolution. Captures mountain channeling and valley funneling (e.g. Adirondack/Hudson valley flows). Goes out to 48 hours into the future.	Less frequent updates. Can get slightly stale on fast-changing front times.
NBM	2.5 km / 10-day range / Calibrated Blend	Most accurate wind speed forecasting due to historical statistical bias-corrections. Best baseline for sustained speeds.	Lacks fine-grid directional trends for specific localized thermal shifts.
ECMWF	9 km / 10-day range / 12-hr updates	Gold standard for multi-day planning (3-5 days). Best for prevailing gradient direction and front tracking.	Coarse 9 km grid misses small lake wind funnels and local thermal breeze mechanics.
GFS	13 km / 16-day range / 6-hr updates	Updates 4x/day. Good broad-brush synoptic wind patterns. Helps build planning confidence if aligned with ECMWF.	Often struggles with frontal timing accuracy and overpredicts gust speeds.
ICON	13 km / 7.5-day range / 6-hr updates	German model, fully independent from GFS/ECMWF. Excellent secondary planning opinion.	Low resolution for small-lake localized thermal breeze effects.
NAM	3-12 km / 3.5-day (84h) range / 6-hr updates	Useful for 1-2 day regional gust trends and frontal boundary locations.	Frequently overpredicts sustained winds and can lag on shift timings.
GEM	15 km / 10-day range / 6-hr updates	Canadian global model. Solid broad-brush backup opinion for planning windows.	Coarse resolution, least sensitive to local convective details.

💡 Model Weights Status & Lifecycle ×

Weather models constantly update, and their short-term local performance changes based on active pressure systems and seasons. To guarantee a premium, accurate consensus, the system tracks weights under these strict rules:

1. Expiration (24-Hour TTL)
Weights older than 24 hours are discarded. The system automatically reverts to default equal weights (16.7% each) to prevent using outdated predictive models.

2. Configuration Drift (Stale Flag)
If the active **Validation Period** or **Start Date** is changed, the weights are marked as *stale*. This alerts you that the calculated accuracy rates mismatch your current session parameters.

3. How to Resolve
Click **Step ⑦: Re-Calculate & Save Weights** to run a fresh multi-model validation. This generates and locks in brand new parameters for your active date range.

💡 Validation Baselines & Desired Results ×

The system evaluates weather model forecasts by comparing their past predictions against a historical baseline. You can toggle between two primary types of baselines:

1. ERA5 Reanalysis (Open-Meteo Archive)
* How it works: Uses ECMWF's global ERA5 reanalysis grid dataset (9 km resolution). This dataset blends millions of global physical readings using advanced physical models to reconstruct historical weather on the coordinate grid. * Desired Result: Generates a highly stable, grid-exact independent baseline. It eliminates the self-referential 0.0 error loop of the forecast model, yielding realistic historical accuracy weights.

2. NOAA Weather Stations (Physical Baselines)
* How it works: Fetches real physical sensor measurements (ASOS/AWOS) directly from the NWS API. * Available Stations: * KSCH (Schenectady Airport): 15 mi SSW. Most stable and reliable. * KGFL (Glens Falls Airport): 20 mi N. Excellent northern wind flow proxy. * 5B2 (Saratoga County Airport): 7 mi WNW. Closest station geographically. * Desired Result: Direct comparison against physical anemometers. Helps identify which model best predicts real-world ground conditions, though observations are subject to airport site geography rather than the lake surface itself.