Module 4 Aws

AW S A c a d e m y M a c h i n e L e a r n i n g F o u n d a t i o n s
Module 4: Introducing Forecasting
© 2020, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Module overview
Sections Demonstrations
1. Forecasting overview • Creating a forecast with Amazon
Forecast
2. Processing time series data
3. Using Amazon Forecast Guided Lab
4. Guided lab • Creating a Forecast with Amazon
Forecast
5. Module wrap-up
Knowledge check
© 2020, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 2
Module objectives
At the end of this module, you should be able to:

• Describe the business problems solved by using Amazon Forecast
• Describe the challenges of working with time series data
• List the steps that are required to create a forecast by using Amazon Forecast
• Use Amazon Forecast to make a prediction
Section 1: Forecasting overview
Overview of forecasting
• Predicting future values that are

based on historical data
• Can be either univariate or
multivariate
• Common patterns – Trending data Seasonal data
• Trends: Patterns that increase,
decrease, or are stagnant
• Seasonal: Pattern that is based on
seasons
• Cyclical: Other repeating patterns
• Irregular: Patterns that might appear Cyclical data Irregular data
to be random
Forecasting use cases
Sales and demand forecasts Inventory projections
Energy consumption Weather forecasts

Section 2: Processing time series data
Time series data
Unit Sales Product #21232
Time series data
is captured in
sequence over time
Related data informs

the time series data—
for example, price or
promotions
Metadata might also be

needed to explain
predictions—for
example, brand name or Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
category
Time and date challenges
Incomplete and varying timestamps yyyy-mm-dd HH:MM:SS Includes time
yyyy-dd-mm Year, day, month
yyyy-mm-dd Year, month, day
yyyy-mm No day
ss:s Seconds
mm-dd No year
UTC, local, and time zones:

Is the time in UTC format?
2020-06-02T13:15:30Z
Time series handling: Missing data
Forward Fill Moving Average
Last known == missing value Avg(previous) == missing value
Backward Fill Interpolation

Next value == missing value Linear
Spline
Danger! Lookahead risk! Polynomial
Note: Zero is sometimes the perfect fill value

Time series handling: Downsampling
yyyy-mm-dd HH:MM:SS yyyy-mm-dd
InvoiceDate Item Quantity

2009-12-01 07:45:00 21232 24 InvoiceDate Item Quantity
2009-12-01 10:06:00 21232 36 Mean/ 2009-12-01 21232 60
2010-12-08 14:21:00 21232 17 Sum 2010-12-08 21232 26
2010-12-08 13:11:00 21232 9 2010-12-09 21232 44
2010-12-09 18:28:00 21232 44
Time series handling: Upsampling
yyyy-mm-dd yyyy-mm-dd HH:MM:SS
InvoiceDate Item Quantity

InvoiceDate Item Quantity 2009-12-01 ??:??:?? 21232 ??
2009-12-01 21232 60
?
2009-12-01 ??:??:?? 21232 ??
2010-12-08 21232 26 2010-12-08 ??:??:?? 21232 ??
2010-12-09 21232 44 2010-12-08 ??:??:?? 21232 ??
2010-12-09 ??:??:?? 21232 44
Reasons to upsample:
• Match different time series
• Irregular time series
• Knowledge of domain
Time series handling: Smoothing data
outliers
Smoothing function
Why are you smoothing? How does smoothing

• Data preparation affect your outcome?
• Visualization • Cleaner data to model
• Model compatibility
• Production improvements
Seasonality
• Seasonality frequency
• Hourly, daily, quarterly, yearly
• Spring, summer, fall, winter
• Major holiday sales, winter
holiday season
• Incorporating holidays
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Time series correlations
This chart is originally from Tyler Vigen: Spurious Correlations

Stationarity, trends, and autocorrelation
• Stationarity
• How stable is the system?
Autocorrelation?
• Does the past inform the future?
• Trends
• Correlation issues
• Autocorrelation
• How points in time are linearly
related
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
• Influences algorithm choice Trend
Using pandas for time series data
• Time-aware index
dataframe['2010-01-04']
dataframe['2010-02':'2010-03']
dataframe['weekday_name'] = dataframe.index.weekday_name
• GroupBy and resampling operations

dataframe.groupby('StockCode')
dataframe.groupby('StockCode').resample('D').sum()
• Autocorrelation
dataframe['Quantity'].autocorr()
Time series algorithms
• Autoregressive Integrated Moving

Average (ARIMA) ARIMA
• DeepAR+
• Exponential Smoothing (ETS)
• Non-Parametric Time Series (NPTS) Prophet DeepAR+
• Prophet Algorithm
NPTS ETS
• Time series data is sequenced
Section 2 key
• Time challenges –
takeaways • Different formats
• Missing data
• Seasonality
• Correlations
• The pandas library offers support for
time series data
• With Amazon Forecast, you can
choose between five algorithms
19 © 2020, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Section 3: Using Amazon Forecast
Amazon Forecast workflow
New data, retraining Deploy
model
Business problem
Generate forecasts
Problem Tune Yes
formulation model
Meets
Collect and Evaluate Feature Select and Evaluate
business
label data data engineering train model model
goal?
Import, evaluate, and clean your data Train a predictor No
Amazon Forecast overview
AWS Management
Historical data Console
Amazon Forecast Predictors
Related data CSV file
Supported domains
• Retail • Web traffic

• Inventory planning • Metrics
• Amazon EC2 capacity • Custom
• Work force
Amazon Forecast
Training Domain-specific
Predictor
data schema
Forecast
Retail forecasting example
• Time series data

• Transactional sales data
• Timestamp, item, quantity
• Metadata
• Category, item color
• Item, metadata
• Related data
• Time series
• In-stock data
• Promotion data
• Timestamp, item, price
Web traffic forecast example

• Webpage ID
• Page views per month
• Timestamp
• Related and metadata
• Page category
• Geographic identifier
Selecting an Amazon Forecast algorithm
You can select from the following

list of algorithms:
• ARIMA Predictor
• DeepAR+
• ETS
• NPTS Model
• Prophet
Algorithm
Evaluating your forecast: Back testing
Predictor accuracy metrics are based on back testing.

Forecast horizon
Start date End date
Training
Testing
Time
Back test window offset

Backtest window 1 Training Testing
Evaluation metrics: wQuantileLoss
• Quantiles determined for 10%, 50%, and 90% quantiles

• wQuantileLoss is the average error for each quantile in a set –
• Works best for models with greater variability in the errors
60
Demand
P10 forecast
40
P50 forecast
P90 forecast 20
-20
Root mean square error (RMSE)
RMSE is the square of the errors.
Actual
Test Results Test result Prediction Deltas
1 2 4 2
Error is the Actual result Prediction
2 4 8 4
30 difference 3 8 12 4
between the 28 4 16 28 12
25 actual
target value and RMSE 6.708204
20 the forecasted
value 16
15
12
10
8 8 Works best for a
5 model where most
4 4
2 errors are of similar
0 size
1 2 3 4
Model accuracy example
Web retailer of shoes wants to predict how often it

will be unable to fill orders for AnyCompany
brand shoes.
Amazon Forecast predicts demand of 1,000 pairs

per month P10 = 880
• P10: 10% of the time, fewer than 880 pairs will be ordered P50 = 1050 Forecast = 1000
• P50: 50% of the time, fewer than 1,050 pairs will be ordered P90 = 1200
• P90: 90% of the time, fewer than 1,200 pairs will be ordered
Demonstration:
Creating a forecast
with Amazon
Forecast
• You can use Amazon Forecast for
Section 3 key time series data
takeaways • Schemas are specific for domains
• Data can include –
• Metadata
• Related data
• Data is split into training and testing
data by accounting for time
• Use RMSE and wQuantileLoss
metrics to evaluate model
Module 4 –
Guided Lab:
Creating a Forecast
with Amazon
Forecast
Module wrap-up
Module summary
In summary, in this module you learned how to:

• Describe the business problems solved by using Amazon Forecast
• Describe the challenges of working with time series data
• List the steps that are required to create a forecast by using Amazon Forecast
• Use Amazon Forecast to make a prediction
Complete the knowledge check
Additional resources
• Amazon Forecast documentation

• Amazon Forecast product page
• How to not use machine learning for time series forecasting
• Time series forecasting principles Amazon Forecast whitepaper
Thank you
© 2021 Amazon Web Services, Inc. or its affiliates. All rights reserved. This work may not be reproduced or redistributed, in whole or in part, without prior written permission from Amazon
Web Services, Inc. Commercial copying, lending, or selling is prohibited. Corrections, feedback, or other questions? Contact us at https://support.aws.amazon.com/#/contacts/aws-training.
All trademarks are the property of their owners.

Module 4 Aws

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Module 4 Aws

Uploaded by

Copyright:

Available Formats

AW S A c a d e m y M a c h i n e L e a r n i n g F o u n d a t i o n s

Module 4: Introducing Forecasting

At the end of this module, you should be able to:

Section 1: Forecasting overview

• Predicting future values that are

Sales and demand forecasts Inventory projections

Energy consumption Weather forecasts

Section 2: Processing time series data

Related data informs

Metadata might also be

Incomplete and varying timestamps yyyy-mm-dd HH:MM:SS Includes time

yyyy-dd-mm Year, day, month

yyyy-mm-dd Year, month, day

UTC, local, and time zones:

Backward Fill Interpolation

Note: Zero is sometimes the perfect fill value

yyyy-mm-dd HH:MM:SS yyyy-mm-dd

InvoiceDate Item Quantity

InvoiceDate Item Quantity

Why are you smoothing? How does smoothing

This chart is originally from Tyler Vigen: Spurious Correlations

• Influences algorithm choice Trend

• GroupBy and resampling operations

• Autoregressive Integrated Moving

Section 3: Using Amazon Forecast

Import, evaluate, and clean your data Train a predictor No

Amazon Forecast Predictors

Related data CSV file

• Retail • Web traffic

• Time series data

• Time series data

You can select from the following

Predictor accuracy metrics are based on back testing.

Back test window offset

Backtest window 2 Training Testing

Backtest window 3 Training Testing

• Quantiles determined for 10%, 50%, and 90% quantiles

Web retailer of shoes wants to predict how often it

Amazon Forecast predicts demand of 1,000 pairs

In summary, in this module you learned how to:

• Amazon Forecast documentation

You might also like