Sei sulla pagina 1di 32

Tools and Techniques to Improve the

Accuracy of Short-term Forecasting


Thomas Brassøe
Stavanger October 22nd 2015
What we will cover

Situation and key demand planning challenges

Introduction to demand sensing

What new possibilities does SAP offer in terms of demand sensing?

What can you do in APO Demand Planning to work with short-term


forecasting and demand sensing?

Case study example

Wrap-up

2
We need the right product, at the right place, at the right time
• Companies need their supply chain to deliver when it comes to satisfying customer demands and optimizing their
inventories
• This task is not easy, particularly in FMCG companies that often need to focus much more on the short-term horizon
– Difficult to integrate real-time data into existing time series statistical model
– Same time series model applied across short-, mid-, and long-term plans
– Time consuming to evaluate statistical models across many planning combinations
• Focusing on short-term forecasting and demand sensing functionality attempts to address these concerns and has
shown to improve the forecast accuracy in the near term

3
Sometimes we see a gap in functionality of creating a good short-term
forecast when we look at the planning hierarchy in relation to SAP

Planning hierarchy Forecast used for Functionality

Strategic
• Annual budgeting • IBP S&OP
planning
• Annual strategic planning • Long-term statistical forecast (business
5 years
group/product line)

Sales and operations


planning
• Sales and operations planning • Mid- to long-term statistical forecast
2-18 months
(product group)
• Demand management
• Sales input to create consensus forecast
• Inventory planning
• Promotion planning
Tactical planning • Master production planning
1-3 months

• Production planning • Forecast consumption


Operational planning • Promotion planning • Short-term statistical (SKU)
0-6 weeks • Call-off on purchase • Demand sensing
• Replenishment and deployment

4
What we will cover

Situation and key demand planning challenges

Introduction to demand sensing

What new possibilities does SAP offer in terms of demand sensing?

What can you do in APO Demand Planning to work with short-term


forecasting and demand sensing?

Case study example

Wrap-up

5
Demand sensing — What does it mean?
• Demand sensing is a next-generation forecasting method that leverages new mathematical techniques and near
real-time information to create an accurate forecast of demand, based on the current realities of the supply chain
• The typical performance of demand sensing systems reduces near-term forecast error by 30% or more compared to
traditional time series forecasting techniques (Source: Wikipedia)

Source: SAP

6
Demand sensing vs. traditional forecasting (Demand Planning)
• Traditionally, forecasting models were based on time series techniques that create a forecast based on prior sales
history
– However, where past sales are great for predicting future forecast levels in mid- and long-term horizons, it has
shown to be less accurate for the short term
• Demand sensing is fundamentally different in that it uses a much broader range of demand signals (including current
data from the supply chain) and different mathematics to create a more accurate forecast that responds to real-world
events such as market shifts, weather changes, natural disasters, consumer buying behavior, etc.
• This lets the planners plan and frees up valuable time so they can focus on longer-term strategic and promotional
activities

7
A simple example of how demand sensing works: Using newest
sales history to update forecast

Demand sensing uses advanced


predictive algorithms to adjust the
forecast in short-term horizon

8
Another simple example of how demand sensing works:
Using bias in promotion to adjust forecast

The bias in the promotion is used by


the demand sensing algorithms to
adjust the forecast down

9
Different sources for predicting the demand-sensed forecast
Internal sources
• Sales orders (historical and future) and shipments
• Final forecast together with forecast accuracy and forecast bias
• Campaigns and promotions
• New product introductions
External sources
• POS data from stores
• Economic trends in the market (e.g. financial downturns)
• Social network sentiment
• Changes in consumer behavior
• Impacts of extreme weather such as disasters

10
What we will cover

Situation and key demand planning challenges

Introduction to demand sensing

What new possibilities does SAP offer in terms of demand sensing?

What can you do in APO Demand Planning to work with short-term


forecasting and demand sensing?

Case study example

Wrap-up

11
SAP components for demand planning with IBP landscape

SAP Supply Chain Control Tower

IBP for sales and operations IBP for response and supply IBP for inventory

Integrated Business Planning for Demand (IBP)


Including Demand Planning and Demand Sensing

Downstream
demand signals
Downstream
demand signals
Master data
Forecast
Delivered order
Social media DSiM adjustments
Open orders

Master data
Delivered order

ERP

12
SAP components for demand planning Adoption scenario
Sales & Operations Planning (IBP)

Forecast S&OP Forecast

APO (Mid-to Long Term Demand Planning)

Downstream Short-term
Forecast
Downstream demand signals forecast
demand signals
Downstream demand signals IBP for Demand
Social media DSiM
sensing*

Master data
Master data Forecast
Delivered order
Delivered order adjustments
Open orders

ERP

* An earlier on-premise version of IBP for Demand sensing was called Enterprise Demand Sensing (EDS)

13
The algorithms in SAP Demand sensing does not only calculates a
better daily forecast pattern it also adjust the weekly forecast

Weekly adjustments Optimal blend by lag


• Downstream data patterns
correlation corrections
• Recent demand
trends and shifts
• Intelligent forecast
consumption logic
• Daily outputs for
Algorithm determines execution systems
optimal weighting with
pattern recognition of • Aggregated weekly or
each input to determine monthly values for
Daily disaggregation
sensed demand planning and reporting
• Forecast pattern
• Pattern prediction
• Shipments patterns
• Back-feed for improved
• Downstream data supply chain planning
pattern

14
What we will cover

Situation and key demand planning challenges

Introduction to demand sensing

What new possibilities does SAP offer in terms of demand sensing?

What can you do in APO Demand Planning to work with short-term


forecasting and demand sensing?

Case study example

Wrap-up

15
What can you do in APO Demand Planning to work with short-term
forecasting and sensing?
• Using the most recent history to calculate a short-range forecast
– “Sensing” using current statistical model
• Calculating daily forecast pattern
– Disaggregate the forecast based on historical patterns instead of average split
• Using incoming future orders to predict changes in demand
– Replace the unknown with the known
– Alerts for big orders
– Customer forecast to indicate customer behavior

16
Using the most recent history to calculate a short-range forecast
• Combining two forecast models to support both short-term and long-term planning horizons
– Weekly short range – e.g., using a constant model (higher alpha)
– Monthly long range – e.g., using a seasonal model (low alpha)
• Combining the two forecast models can be done by using a macro

17
The daily forecast pattern
• As an alternative to splitting days equally, history is used to create a daily forecast pattern with the aim to achieve a
better daily forecast
• The logic processes history weeks and averages each day’s proportion into the Daily Forecast Pattern

• Either done by using standard macro functionality or based on user exit macro

18
Degrees of uncertainty — as soon as possible, replace the
unknown with the known
• A forecast is just an order we have not got it yet
– Replace the forecast with something more certain; ideally, customer orders if possible
» Quotes are less uncertain than forecast
» Commitments from customers on, e.g., promotions are also less unknown, less uncertain

The past Today The future

HIGH Degree of demand certainty LOW

Facts Known Partially known Unknown

Deliveries Orders Commitments Quotes Forecast


Source: Model based on Tom F. Wallace, Sales Forecasting – A new Approach

19
Use macros to generate alerts if incoming orders exceed the
created forecast outside order horizon
W 05.2014 W 06.2014 W 07.2014 W 08.2014 W 09.2014 W 10.2014 W 11.2014

Color codes are used to reflect the ratio


Committed orders are displayed in between “forecast” and “order” in
the planning book. specified horizons. Alerts are created
for exceptions.

20
Using macro functionality to replace the forecast with the
committed orders
W 05.2014 W 06.2014 W 07.2014 W 08.2014 W 09.2014 W 10.2014 W 11.2014

Committed orders are After the order horizon, the


displayed in the planning book. unconstrained forecast is
replaced by the committed
orders.

Whether the forecast should be changed or not depends on the lead time in production.

21
Customer forecast to indicate customer behavior

The built-in demand sensing is built on the


customer forecast input.
• The first check is whether the customer
forecast has changed since yesterday. If the
changes are significant (+10%), a color code
appears in the date view, and an alert is
created
• If the customer forecast is significantly different
from our forecast (+20%), an alert is created

22
Even though APO is not really demand sensing software,
there is still much you can do
• Running a short-term forecast and combining this with the long-term forecast will make it possible to utilize near-term
information
• It is possible to work with daily buckets in APO Demand Planning and thereby split the weekly forecast in other ways
than equal
– It can impact performance though
• Incoming orders and other demand streams can be used to sense signals in the near-term demand pattern
– Alerts can be used to guide the planners to update the forecast either manually or automatically via macros

23
What we will cover

Situation and key demand planning challenges

Introduction to demand sensing

What new possibilities does SAP offer in terms of demand sensing?

What can you do in APO Demand Planning to work with short-term


forecasting and demand sensing?

Case study example

Wrap-up

24
Case example of a proof of concept for an FMCG company
Situation
• Company relies heavily on accurate daily forecasts to optimize replenishment stocks whilst ensuring a close-to 100%
service level. Shelf life, in some instances, is <7 days
Complication
• The amount of computations to create a daily short-term forecast is quite substantial in terms of data volume and
there are several developments and customizations in APO DP to support the process. In addition, much effort from
demand and supply planners goes into tuning and adjusting the daily short-term forecast
Main question is:
• Can Demand Sensing (EDS) outperform, or match, the current short-term daily forecast (including manual planner
inputs) in terms of forecast accuracy?

Timeline

Data extract and


Test data extract Scope determination FCA benchmark
analysis

Determine data availability Finding business Submitting data and Benchmark to current
and manual data extract scenario EDS algorithm run forecasts
efforts representatives, in
terms of products and
dates

25
Proof of concept — Environment
The proof of concept runs on three major overall considerations:

“Base period”: “Base period”:


• From which date (in history), is the split between • As many products face seasonality because of
what should be regarded as history, and what should public holidays, the base period was decided to be
be regarded forecast in the PoC? just before of Easter 2014.
Forecast granularity: Forecast granularity:
• In which time, resolution will the forecast data be • As the purpose of EDS is to create a short-term
delivered? daily forecast from the long-term weekly forecast,
the input to EDS will be in weekly buckets.
Sales order granularity: Sales order granularity:
• In which time, resolution will the sales order and • As for generating daily pattern, EDS needs the sales
shipment data be delivered? order granularity to be daily.

Source: SAP

26
Proof of concept — Data requirements and manual data processing
In addition to master data for product (item), location, and customer (ship-to location), EDS utilizes three transactional data
inputs:

What to be aware of: EDS


• Ship-to location ID is usually
not a part of the planning
Long-term forecast Sales orders Shipments (/deliveries)
objects in APO (a part of the
CVC structure). Hence, sales
orders must be aggregated Item ID Order ID Sale shipment ID
and filtered manually,
according to BW logic to Location ID Sale order line ID Order ID
reflect the customer dimension
in APO DP Ship-to location ID Item ID Sale order line ID
• Because of the above,
discrepancies between sales Revision date Location ID Shipment date
order lines and BW history in
APO can occur
Forecast date Ship-to location ID Shipment quantity

Quantity Creation date

Requested delivery date

Requested quantity

Open quantity

27
Delivering forecast versions and current forecast setup
Current forecast buildup:

Several, hard-to-forecast, factors


Promotion influence the unconstrained forecast:
forecast Planner forecast • Size and frequency of promotions
Statistical
(If any) (if any) Unconstrained
forecast without • Public holidays
forecast
promotions
• Sales and planner knowledge of
customer behavior
Input sequence

• The long-term unconstrained forecast is supplied to EDS


• The forecast is delivered in multiple revisions throughout the
assessment period. BW picks up the forecast weekly, hence
weekly revisions of the forecast are provided.

28
Benchmarking and result
• In total, 1.8m records were delivered to EDS, 279 CVCs across products groups, customers, and locations
• Two kinds of forecast error benchmarks are available – in weekly and daily buckets

Baseline: Long-term forecast1 Baseline: Long-term forecast equally split into days1

• Baseline error across all product groups: 39% • Baseline error across all product groups: 70%
• EDS error across all product groups: 32% • EDS error across all product groups: 51%
Weekly

Daily
Gain: 18% Gain: 27%

Baseline: Current short-term forecast based on


customizations and developments2

• Baseline error across all product groups: 56%


• EDS error across all product groups: 69%
Daily

Loss: 24%

Of 131 of the CVCs3 used in the period, EDS provided


better forecast on the CVC level for 21 of them

1 Based on the assessment period (51 weeks)


2 Based on the forecast period (3 weeks), including a public holiday (Easter)
3 Not all 279 combinations have sufficient history or forecast in the 3-week benchmark period

29
Conclusion and things to be aware of
• EDS has shown to improve the weekly forecast in near-term horizon and is significantly better on daily level
compared to an average split
• In the Easter week where planners are spending lots of time getting the final forecast right, the EDS does not
outperform the human touch
– Be careful in choosing scope and measuring periods
– Measure also more “normal” weeks to see effect
• In general, it has shown that products without manual planner attention are performing better in EDS than products
with lots of planner attention
– Using EDS as the first step before planner input could be a way to utilize the best of both “worlds”

30
What we will cover

Situation and key demand planning challenges

Introduction to demand sensing

What new possibilities does SAP offer in terms of demand sensing?

What can you do in APO Demand Planning to work with short-term


forecasting and demand sensing?

Case study example

Wrap-up

31
Your turn!

How to contact me:


Thomas Brassøe
thb@implement.eu

32

Potrebbero piacerti anche