Amazon Forecast Overview

ADVANCED MACHINE LEARNING TIME-SERIES FORECASTING

Thi Nguyen
Solutions Architect

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
The power of forecasting
IMPROVING BUSINESS OUTCOMES WITH MACHINE LEARNING

INVENTORY PLANNING WORKFORCE PLANNING

Improve demand planning at granular levels More effectively staff to meet

varying demand levels
Reduce waste, increase inventory turns,
and improve in-stock availability Improve utilization, time to serve,
and customer satisfaction

CAPACITY PLANNING FINANCIAL PLANNING

Make longer term decisions with more confidence Plan for sales and top-line revenue
Improve capital utilization Effectively manage cash flows

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
The current landscape
Customers increasingly want more variety, they
want it immediately, and they want it cheaply

Accurate and flexible forecasts are key; however,

traditional methods cannot capture the increasing
number of demand signals and complexity

Market leaders are investing in ML-driven

forecasting to more effectively meet demand

20% accuracy improvements

5% inventory reduction
3% increase in revenue

Harvard Business Review, McKinsey Global Institute

https://www.mckinsey.com/business-functions/mckinsey-analytics/our-
insights/most-of-ais-business-uses-will-be-in-two-areas

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Foxconn saves $500K annually
MANUFACTURES SOME OF THE MOST WIDELY USED ELECTRONICS WORLDWIDE

Challenge Solution
● Limited data science experience ● Built a custom forecasting

internally for Mexico factory. solution with Forecast in 2

● Having individual forecasts for months.
● Removed known anomalies.
each product is important to
understand the mix of skills ● Used AutoML in Forecast to

needed in workforce. overcome limited background

in time-series modeling.

© 2022, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Meesho increased forecasting accuracy by 20%
INDIA'S LARGEST MARKETPLACE FOR 100 MILLION SMALL BUSINESSES

Challenge Solution
● Out of stock (OOS) products ● Frequent model retraining to
degraded customer experience. capture trend and seasonality.
● Suppliers had no visibility on ● Weekly forecast of inventory
inventory needs. for suppliers.
● Leveraged probabilistic
forecast using quantiles.

© 2022, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Amazon’s evolutionary forecasting journey
STATISTICAL METHODS ML METHODS ML + DEEP LEARNING

1995 2000 2005 2010 2015 2020

Best for few or independent Best for complex and inter-related data
data (e.g., aggregate demand) (e.g., thousands of SKUs at hundreds of stores)
• Aggregate demand • Regional vs national demand
• Everyday household • Subtle seasonal patterns
products • Relationships between products
• Basic trends • New products
• Repeated seasonal patterns • Slow moving products
• Changing prices or promotions
• External events
© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved. 6
 Highly accurate Train and deploy ML models which
can be up to 50% more accurate than traditional
methods. No machine learning experience required

No code and no machine learning experience

required. Leverage all the heavy lifting of building,
training, and deploying custom models at scale
Leverage LCNC ML to
optimize service levels
Easy to integrate into existing data lake, inventory,
Amazon Forecast ordering, and supply chain systems
Amazon SageMaker Canvas

Quickly iterate, explore, prepare data, and onboard

for ML-based forecasting. No need to bring entire
data architecture onto AWS.

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Purpose-built low code no code (LCNC) services

Business Data
Developers Scientists
Analysts
Code to Amazon Use existing
Drag-and-drop Forecast APIs, Amazon SageMaker
forecasting using integrate into JumpStart models
Amazon SageMaker existing systems or build custom
Canvas
models

- Select levels of AutoML - Custom cases where current

- Visual, drag and drop interface
- S3 data ingress and egress algorithms don’t meet needs
- Explore data
- MLOps using AWS Step Functions - Existing solutions in
- Include historical, related, and
- Detailed backtest results JumpStart or BYO custom
item metadata data
- Choice of quantiles, algorithms container to SageMaker
- Preview, Quick, and Standard
- Automatic model monitoring - Full control over algorithms,
(Amazon Forecast) Models
MLOps, inference

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Fully managed ML forecasting
PREPARE, BUILD, TRAIN, TUNE, DEPLOY, AND MANAGE

AWS Low Code No Code Forecasting

Data preparation Train and tune Deploy and manage

Inspect Feature- Built-in Train/ Select & Train & Optimize Calculate Host Compute Predictor
data; specific data sets test tune hyper- optimize ensembled accuracy trained & host monitoring
fill imputation (weather/ split parameters multiple model metrics, models inference
missing holidays) base for each explainability
rows models time series impact scores

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Forecasting workflow
AUTOMATED FORECASTING BY SCHEDULES OR BY EVENTS

Integrate &
build consensus
on predicted data

Extract latest Import Generate

data input(s) data set(s) forecasted
data points

Monitor
model performance

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Forecasting inputs and outputs

Historical data
Sales, inventory, call Amazon Forecast
volume, resource demand SageMaker Canvas

Related data
Price, promotions, in-stock,
Forecasts Explainability
weather, custom events

Item metadata Built-in datasets

Color, city, category, brand, (holidays, weather)
author, size

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Data input examples
Target time series Related time series Item metadata
(Historical demand)

Timestamp Item Id Location Price Item Id Color Category

Timestamp Item Id Location Target value
Jan 1 1111 US_98121 $5 1111 Yellow Outdoor
Jan 1 1111 US_98121 50
Jan 1 2121 US_98121 $20 2121 Red Outdoor
Jan 1 2121 US_98121 89
Jan 2 1111 US_98003 $5 3434 Green Indoor
Jan 2 1111 US_98003 35
Jan 2 3434 US_98003 $13 5000 Blue Indoor
Jan 2 3434 US_98003 73
Jan 3 1111 US_98003 $5 Category, department, color, texture
Jan 3 1111 US_98003 45
Jan 3 5000 US_43505 $4
Jan 3 5000 US_43505 13
Price, promotion, events, store hours,
Demand, sales, call volume, energy consumption, cloud competitive average price
storage consumption

Tips:
1. Identify data sources and validate data is time series.
2. Gather large number of item IDs and historic data points (e.g. 3 years of daily data).
3. Clean and prepare data in consistent time stamp and frequency.

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
No code forecasting with SageMaker Canvas
POINT AND CLICK INTERFACE TO QUICKLY GENERATE MACHINE LEARNING FORECASTS

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Single-item or batch forecasts

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Generate ‘what-if’ scenarios
• What is the impact on sales if
I decrease prices?
• How do external factors
impact demand?

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Integrate Amazon Forecast into your process
AWS STEP FUNCTIONS SEQUENCES AMAZON FORECAST API CALLS

Amazon Forecast

Amazon S3 Amazon Forecast Amazon Forecast Amazon Forecast Amazon S3

Athena connector
(model training data) Import data Model training Produce predictions (results)

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Understand your forecasts with explainability

Explainability Attribute increasing impact score

Base price 0.87

• Quantify relative impact of
each attribute Emailer for promotion 0.11

• Directional impact (+/–) of Checkout price 0.11

each attribute 0 0.2 0.4 0.6 0.8 1

• Model or time-series level Attribute decreasing impact score

explanations Food category 0.61

• Supports drill down to Food cuisine 0.34

specific time points for time
series Homepage featured 0.08

0 0.2 0.4 0.6 0.8 1

© 2022, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Balancing under-forecasting and over-forecasting
FORECAST AT VARIOUS QUANTILES OF THE PROBABILITY DISTRIBUTION

P50 P75

50% 75%
probability probability of
of meeting meeting demand
demand

100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000

Demand (units) Demand (units)

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Forecasting customers

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
The AWS Prototyping Program helps customers contextualize
complex uses of the AWS Platform and associated services by
building working prototypes of specific customer use cases,
leveraging their existing data, devices and systems.

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
ASEAN Prototyping Assets
Visual Product Offers/Products Omni-channel
recognition personalization engagement platform

Built in 4 weeks, Built in 6 weeks,

Built in 6 weeks, leveraging Amazon SageMaker &
leveraging AWS Rekognition leveraging Amazon Personalize Amazon Pinpoint

Last mile delivery Shelf Monitoring & eKYC

refurbishment

Built in 12 weeks, Built in 6 weeks,

Built in 4 weeks,
leveraging Amazon Event Bridge, leveraging AWS Rekognition
leveraging Amazon SageMaker
Amazon Kinesis and IoT Core
21
© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.

+ Amazon Forecast pipeline, IoT Hub…

Architecture for an Forecast

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Further resources
Main Page: Canvas | Forecast
Pricing: Canvas | Forecast
Documentation: Canvas | Forecast
Customer Use Cases: Canvas | Forecast
Blogs: Canvas | Forecast
GitHub / MLOps: Forecast
Contact your AWS account manager or email [email protected]

© 2024, Amazon Web Services, Inc. or its affiliates. All rights reserved.