Overview of The Weather Companys Principal Forecasting
Methodologies
Peter Neilley, Bruce Rose, Joseph Koval, Todd Hutchinson, Paul
Bayer, Jeff McDonald, John Mathews, William Cassanova, Dale Eck,
Neil McGillis, Shaun Tanner The Weather Company Eric Floehr
Intellovations, LLC or How Atmospheric and Computer Sciences have
created Forecasts on Demand NWA 2015, Plenary III Big Data and the
Traditional Forecasting Paradigm
Forecast foundational datasets are exploding Weather Obs Models
Final Forecast Users Post Procs Forecasters which are overwhelming
the system NWA Annual 2015, Neilley et al. TWCs New Forecasting
Paradigm
Therefore, we set out to create a new forecasting paradigm. Some
key tenants of our approach were to: Take fully-appropriate
advantage of the richness of modern NWP output Allow forecasts to
update continuously at the pace of the input data Preserve human
forecaster influence Be easily adaptable to new input data and new
scientific methods Have globally ubiquitous content Optimize
forecast accuracy NWA Annual 2015, Neilley et al. Some Key
Characteristics
The Weather Companys Forecasts on Demand Paradigm Some Key
Characteristics On Demand: Forecastsare created and deliveredat the
time of request Human Input: Forecasterinfluence retained Fresh:
Forecasts alwaysbased partly on the latestfrom all input sources.
Precise: Forecasts builtfrom full resolution inputdata, not
interpolated pointforecasts. Optimized:Variousstatistical and
scientificmethods govern optimalforecast assembly Forecasters
Forecasts On Demand Users NWP Post Processors A true On Demand
forecast creation system NWA Annual 2015, Neilley et al. Real-time
Publication and Distribution Proprietary Weather Obs
The TWC Forecasts on Demand EcoSystem Obs on Demand Engine
Real-time Publication and Distribution On Demand Users Govment
Weather Obs Proprietary Weather Obs Govment NWP Models Proprietary
NWP Global Weather Data Forecasts on NWA Annual 2015, Neilley et
al. Forecasts on Demand Engine Human Forecasters Over the
Loop
The Forecasts on Demand Engine Forecasts on Demand Engine For
details see: AMS WAF2015: 12B.8and AMS16 EIPS: 13.B1 AMS
WAF2015:Hutchinson et al., 12B.7 AMS WAF2015: Rose et al., 7B.5
Input Weather Data 1-15 Day Forecast Engine Forecasts on Demand
Core Users 0-6 hr Forecast Engine Human Forecasters Over the Loop
NWA Annual 2015, Neilley et al. TWCs Forecasts-on-Demand 1-15 Day
Forecast Engine:
A multi-model, multi-method blending approach 162 inputs,
including: Global and regional models Individual ensemble
members(where available) Various MOS forecast products
Reforecast-based post- processor outputs. A statistically-optimized
blendis based on recent verificationand continuously updated The
blend is unique at eachlocation, time, parameter System updates
with eachnew input arrival For details see:AMS WAF2015: 12B.8Koval
et al., NWA Annual 2015, Neilley et al. TWCs Forecasts-on-Demand
0-hr Forecast Engine:
A blended nowcasting approach Key Characteristics Multi-method,
nowcasting system using 6 inputs Principal inputs to the
precipitation forecastsare: Radar advection forecasts (where
available) Our internal, HRRR-like global NWP Human forecasters.
Forecasts update as new input data arrives, no less than once every
15 minutes (in No. Amer and Europe) See AMS WAF201512B.7Hutchinson
et al. NWA Annual 2015, Neilley et al. Percent Correct
Forecasts
Comparative Forecast Skill: YTD - US Day 1-3 Source:
ForecastWatch.com From ForecastWatch 2015 YTD Overviewstatistic,
defined as: Overview = .25 * MaxT %Cor + .25 * MinT %Cor+ .50 *
Precip %Cor Computed using about 340K forecasts from each providor
from 770 1st order US sites Not shown here: Redundant TWC brands
Redundant NDFD Climo & Persistence Percent Correct Forecasts
NWA Annual 2015, Neilley et al. US, Precip ETS (>0) Days
1-3
Broader Accuracy Metrics Source: ForecastWatch.com Max Temp Min
Temp Precipitation US, MaxT MAE Days 1-3 US, MinT MAE Days 1-3 US,
Precip ETS (>0) Days 1-3 MAE (F) MAE (F) ETS US, MaxT %Busts
Days 1-3 US, MinT %Busts Days 1-3 US, Precip %Cor Days 1-3 Percent
Percent Percent NWA Annual 2015, Neilley et al. Overview statistic:
Combined Mx/Mn/Pcp %Correct
Global Metrics: Overview statistic:Combined Mx/Mn/Pcp %Correct
EUROPE US ASIA Day 1-3 Days 6-9 Days 3-5 Source: ForecastWatch.com
Forecasts on Demand Operations
Deployed in the Amazon cloud in 4 global regions Typically delivers
11B forecasts/day Peak load of 26B/day or 250K/sec. Mean forecast
creation time ~ 11 ms. Mean total request-to-delivery time < 300
ms. Used by all TWC/WSI consumer forecast systems (The Weather
Channel, weather.com, WeatherUnderground, Intellicast, etc.) Also
drives our partners weather including Apple, Google, Yahoo!, IBM
and a majority of the domestic TV stations. NWA Annual 2015,
Neilley et al. Moving Forward Expansion of input data, particularly
internationally
Models, (e.g. UKMO, GEM, JMA, FIM global models), Additional
observations and remote sensing for nowcasting Non-linear
multi-model integration techniques Forecast PDFs and
event/threshold forecasts (e.g. prob,.of 6 of snow) Weather object-
based blending techniques NWA Annual 2015, Neilley et al. From the
AMS Weather Anal. and Forecasting Conf, 2015:
For More Details...... From the AMS Weather Anal. and Forecasting
Conf, 2015: 7B.5 A Human Over the Loop Forecast Paradigm at The
Weather Company (TWC), Rose et al., 12B Day Weather Forecast
Guidance at The Weather Company, Koval et al. 12B hour Weather
Forecast Guidance at The Weather Company, Hutchinson et al Upcoming
at the AMS Annual Meeting: EIPS 13B.1 Recent Advances in The
Weather Companys 1-15 Day Forecasting Guidance Infrastructure,
Koval et al. ProbStats 6.5 Consensus forecasting using constrained,
regularized regression. Williams et al. NWA Annual 2015, Neilley et
al.