32
1 Powerful forecasting for a good planning VisionWorks Seminar 25/02/2014

Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

  • Upload
    ordina

  • View
    207

  • Download
    2

Embed Size (px)

DESCRIPTION

 

Citation preview

Page 1: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

1

Powerful forecasting for a good planningVisionWorks Seminar 25/02/2014

Page 2: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

2

Contents

Forecasting and planning – a perfect interplay

What to forecast and how to forecast it

Forecasting with Ordina

- The bForecasting case

- The Pluto Forecasting case

Page 3: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

3

Forecasting and planning – a perfect interplay

Page 4: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

4

What is forecasting?

From businessdictionary.com:

- A planning tool that helps management in its attempts to cope with the

uncertainty of the future, relying mainly on data from the past and present

and analysis of trends.

Data from the past

Trends

Data from the present Certainty of the future?

Page 5: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

5

Forecasting for planning ends

Forecasting becomes useful in a planning context as soon as the important

planning decisions must be based on

- Need for a certain product, e.g. the need for certain consumer goods such as beer,

canned goods, …

- Need for a certain service such as airport security, roadside assistance, shipment

transportation, ...

An accurate forecast leads to a good mid term and long term (capacity)

planning.

A good mid term and long term planning leads to a good short term planning.

This leads to cost reduction as well as customer satisfaction:

- No external parties need to be used to reach SLA

- Capacity is available to ensure in time delivery

- Stocks can be maintained at optimal levels

- …

Page 6: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

6

What to forecast and how to forecast it

Page 7: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

7

Before deciding to use forecasting.

When considering forecasting to have a substantiated basis for long term

planning, we need to answer several questions.

1. What planning decisions do we want to make and what do we base these

decisions on?

2. What are the main factors that influence the basis for these decisions?

3. At what level of detail can we make a prediction?

4. Can we refine the prediction as we process in time?

An answer to these questions will

- not only tell us what to forecast,

- but also what techniques we should use to create this forecast.

Page 8: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

8

What are the main factors that influence the basis of these decisions?

- Historical trends of incidents and B2B agreements.

- Historical trends of visitors, B2B agreements and commercial campaigns,

new product and service launch, …

- Historical trends and customer announcements.

What planning decisions do we want to make?

- E.g. roadside assistance: we want to minimize the use of external parties

needed to maintain our customer service levels.

- E.g. airport services: we want to optimize our time to service and minimize

our personnel cost while maintaining our customer service levels.

- E.g. postal services: we want to optimize machine utilization and minimize

personnel cost while maintaining our target throughput times.

Some examples

Based on the number of incidents on the road.

Based on the number of visitors and passengers at the airport.

Based on the postal volumes received each day.

Information and

knowledge from

other divisions

Historical trends Short term

operational

information

Page 9: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

9

Different forecasting methodologies

Consensus forecasting

- Several parties each make a separate forecast, based on their experience

and knowledge.

- These separate forecasts are combined together to form a final forecast.

Statistical forecasting

- Mathematical techniques are used to extrapolate historical data to the

future to form a final forecast.

Combining forecasts

- Forecasts created using different techniques are combined to form a final

forecast.

- Typically, a statistical forecast serves as the basis for the forecast. It is

subsequently enriched with information received from other channels to

form a final forecast.

Gartner (september 2012)

Defining the balance between statistical modelling and collaborative forecasting

improves accountability for the forecast, and enables continuous improvement

across the organizationCompanies can benefit from clearly defining the balance between statistical modelling and

collaborative forecasting methods to improve accountability for the forecast and put in place

continuous improvement plans to improve the forecast. […]

Page 10: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

10

Good forecasting uses the best of all worlds

Advanced statistical

techniques

Relevant forecast information

from all divisions

Last minute operational

informationActuals

Weather forecast

Historical data

B2B agreements

Sales campaigns

Experience

Page 11: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

11

Forecasting with Ordina: 2 cases

Page 12: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

12

Two Forecasting Cases

bForecasting: the bpost Volume Forecasting Tool

Pluto Forecasting Tool: volume forecasting for roadside assistance.

Both cases were modelled using the Quintiq Software Suite, specifically

designed for modelling Advanced Planning and Scheduling software.

Page 13: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

13

bForecasting – postal Volume Forecasting

Page 14: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

14

o Within its Vision 2020 business plan, increasing the efficiency of its Industrial Mail Centers is an absolute necessity for bpost.

o For an efficient planning, accurate predictions of future mail volumes are necessary, which means 8 different dimensions need to be taken into account

o Moreover, dynamic corrections of the predicted volumes with newly received data must guarantee estimate accuracy up to the hour of the execution of the actual planning.

o Finally, future changes in the bulk of mail volumes received leads to the necessity of having a dynamic identification of relevant statistical dimensions and corresponding breakdown layers.

o bForecasting allows the user to create forecasts with time series of 8 dimensions, where the statistical dimensions can be changed dynamically over time.

o User-extendable advanced statistical algorithms allow full flexibility in statistical forecasts, while dynamic allocation of statistical dimensions ensures future predictability.

o A dynamic breakdown management ensures good predictions up to the most detailed operational level.

o Advanced enrichments with operational data is possible up to the last minute, ensuring operational correctness of the predicted volumes.

services

The bpost Forecasting Case

SO

LU

TIO

NA

DV

AN

TA

GE

S

Page 15: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

15

Bpost: be the strongest and most trusted postal operator

• Bpost is the leading postal operator of the country, with is Mail Service

Operations (MSO) achieving 94% on time delivery.

• The efficient working of its Mail Service Operations (MSO) is crucial for

maintaining its position as strongest and most trusted mail operator in

the rapidly evolving market of postal services.

• An important factor in the bpost delivery process is the sorting which

takes place in its five Industrial Mail Centers: Bruxelles X, Antwerpen X,

Gent X, Charleroi X and Liège.

Page 16: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

16

Bpost: planning the industrial mail centers

The five industrial mail centers are responsible for sorting mail and

parcels received in their regions.

The sorted mail and parcels are subsequently transported to the

regional centers for final sorting and distribution.

To sort the mail and parcels received from the various intake channels,

a large number of sorting machines and their operators need to be

planned.

For the planning to be efficient, accurate predictions are needed so

as to reserve the necessary resources in time and to ensure optimal

usage of machine and personnel capacity.

Page 17: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

17

Bpost: properties of volumes to be sorted

The volumes that need sorting depend on eight different dimensions.

- Intake channel: through which channel are the volumes collected (from the

red letter boxes, from customer drop offs, from the foreign mail centers, …)

- Customer: the corporate customer, if any, dropped the volume.

- Day Plus: how many days after being collected should the volume be

delivered?

- Intake location: where was the volume collected?

- First Sorting IMC: which industrial mail center executes the first sorting

step?

- Mechanization level: can the volume be sorted automatically or will it need

manual sorting steps?

- Throughput type: the size and type of the volume (normal size envelopes,

large size envelopes, parcels, …)

- Sorting level: the extend to which the volume was already sorted.

5

15,000

12

550

2

8

8

3

Page 18: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

18

Bpost: first step towards predictability

For data to have any chance of being predictable, enough volumes

should be known so as to discern patterns in the data.

For the dimensions customers and intake location, a large number of

single volumes exist.

To have any chance at predicting volumes, these single volumes need

to be regrouped.

For this reason, customer pools and location pools were introduced.

Complexity is added as these pools depend on other dimensions and

vary through time.

- E.g. the customer pool for D+1 volume might be different from the pool for

D+2 volume.

- E.g. the customer pool for D+1 in September might be different from the

one in November.

Page 19: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

19

Bpost: more steps towards predictability

Not all dimensions have statistical significance. An important exercise

is to identify those dimensions that have.

Other dimensions need to be derived from the statistical ones using a

breakdown hierarchy and breakdown factors.

As with pools, both hierarchy and breakdown factors depend on other

dimension values as well as on certain time periods.

A large number of breakdown algorithms is available for the

computation.

Page 20: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

20

Bpost: time series for forecasting

Having correctly regrouped dimensions in statistical and operational

dimensions, bForecasting creates all time series containing historical

data.

These time series are then used to predict the future volumes using

advanced statistical algorithms.

bForecasting uses R as an underlying statistical engine, offering all

power and flexibility of the de facto open source standard in statistical

computation.

Page 21: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

21

Bpost: refining the forecast

From its B2B and B2C customers, bpost typically receives detailed

information on the volumes that will be dropped at the MassPost intake

locations one week in advance.

This information is processed by bForecasting an replaces – except

when user-overridden – the statistically computed volumes with the

new information

Page 22: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

22

Bpost: operational follow-up

From operations, bForecasting receives hourly updates on volumes

actually dropped at the IMC.

These volumes are processes using bpost-defined consumption logic

to adapt predictions for the following hours.

Different types of logic can be defined and

assigned to sets of dimension values, giving

full flexibility to the user in predicting the

following hours.

Hourly communication from bForecasting

to the planning tool allows the planning to

be adapted last minute to the volumes

expected in the coming hours.

Page 23: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

23

Pluto Forecasting – Roadside assistance forecasting

Page 24: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

24

Roadside assistance planning

Roadside assistance service providers are highly competitive and need to keep

a competitive edge by increasing their service levels to their members.

Guaranteeing service within 30 minutes, independent of the location of the

incident, can be solved in two ways:

- Position more than enough patrolmen to ensure coverage of the whole of Belgium

- Position just enough patrolmen to ensure coverage of the right areas.

Obviously the first solution is expensive as it introduces a lot of idle time for the

individual patrolmen.

The second solution, however, needs an accurate prediction of the number of

incidents and their geographic distribution.

For this purpose, only an advanced forecasting tool will do!

Page 25: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

25

Pluto Forecasting Tool: forecasting for roadside assistance

For one of the major players on the Belgian roadside assistance

market, Ordina developed the Pluto Forecasting Tool.

Page 26: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

26

Forecasting the volumes – dimensions

The Pluto forecasting tool allows advanced statistical forecasting and

forecast enrichment

- per incident type

- per geographic location (up to address level)

- up to half hour detail level.

Page 27: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

27

Forecasting the volumes – data cleansing

Using the de facto open source standard for statistical computing – R –

a user-extendable number of statistical algorithms is provided for data

cleansing.

Data cleansing in roadside assistance is necessary to eliminate the

inherently unpredictable peaks due to unexpected winter weather or

public holidays.

Season of the

outlier

Percentage

deviation from

historical value

Average deviation:

possible value for the

event’s correction

percentage

Page 28: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

28

Forecasting the volumes – geographic breakdown

For a forecast to be accurate, enough data needs to be available for a

pattern to emerge.

For this reason, the end user can select the geographic level and time

granularity for which statistical forecasts should be made.

Lower level forecasts are

computed using breakdown

factors

- both in time and

- In geographical detail

Page 29: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

29

Forecasting the volumes – statistical forecasting and enrichment

Using R algorithms the user computes and compares forecasts to

arrive at the most accurate prediction for the next year.

Using additional information retrieved from the outlier cleansing, this

forecast can be enriched to model the effects of public holidays and

expected bad weather.

Page 30: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

30

Translating the forecast: occupation requirements

Having created an accurate forecast, this needs to be converted in a

number of shifts that need to be planned in order to achieve the SLA

towards the members on one hand while maximizing the productive

time of the patrolmen on the other hand.

This computation is done in the Pluto Forecasting tool using a greedy

heuristic.

Page 31: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

31

Forecasting in production: surprising results

Roadside assistance incidents prove to be highly predictable on a daily

level:

- Forecast accuracy of over 90%

Moreover, using the Pluto Forecasting tool, long standing “gut feeling”

common knowledge was shown to be wrong:

- “In the summer, we have significantly less incidents than throughout the

rest of the year”.

This claim was shown to be wrong for the patrolmen and right for the call

center and back office.

- “In the winter, we have significantly more incidents than throughout the rest

of the year”.

This claim was shown to be wrong for most of winter, barring the first

couple of days of a cold spell.

Page 32: Ordina - VisionWorks Seminar: Bi Innovation Radar Part2

32

Questions?