Don't Trust the Number Alone: How to Read Visibility Forecasts Like a Pro

2026-03-16

Our site provides AI-powered (LightGBM) visibility forecasts, but taking the predicted value at face value is risky. Like weather forecasts, there's always an error margin. The key is knowing how much error to expect. This article explains forecast accuracy by site and how to read prediction intervals (q10/q90).

Reading Forecasts: Point Estimate vs. Prediction Interval

Example forecast: IOP

12m

Range: 9m15m

→ Read as: 'Around 12m, but likely to fall between 9-15m'

q50

Point estimate (median): most likely visibility value

q10

Lower bound: about 10% chance of being worse. A worst-case estimate

q90

Upper bound: about 10% chance of being better. A best-case estimate

Forecast Accuracy by Site

SiteMAEReliability
Kashiwajima±1.5m0.52High
IOP±2.5m0.82High
Hirasawa±2m0.59Useful
Futo±2.8m0.43Reference
Kushimoto±3m0.43Reference
Yonaguni±5m0.15Use historical avg
Osezaki Bay±4.5m0.20Use historical avg

MAE = Mean Absolute Error (average deviation from actual). R² = coefficient of determination (closer to 1 = better).

Where AI Struggles and What to Do Instead

Yonaguni: Too consistently clear for AI to add value

With a 24.5m annual average and low variance, AI prediction adds little value. Simply expect 20m+ anytime. Use historical averages as your guide.

Osezaki Bay: Dominated by local factors

The enclosed bay geography makes prediction from open-ocean weather data difficult. Local tidal movements and freshwater inflow dominate, so the dive shop's same-day report is most reliable.

Practical Guide to Using Forecasts

When to trust

  • High R² sites: IOP, Kashiwajima, Hirasawa
  • When q10-q90 range is narrow (within ±3m)
  • Under normal weather conditions

When to be cautious

  • After typhoon passage (insufficient data)
  • Wide q10-q90 range (±5m or more)
  • Sites with R² below 0.3

Rule of Thumb: Interpreting Forecast Values

Forecast 12m → Expect 9-15m range

Forecast 8m → Expect 5-11m range

Forecast 20m → Expect 17-23m range

As a general rule, expect the forecast value ±3m. If q10/q90 intervals are displayed, use those instead.

About the Data

Accuracy metrics are from LightGBM model evaluation on test data. MAE is the mean absolute difference between predictions and observations. q10/q90 intervals are from quantile regression. The model uses weather, marine, and satellite data as features.

🌊 Check Visibility Forecasts

View AI-powered 7-day visibility forecasts for 30+ dive sites across Japan.

Open Forecast App →