using neuraltools to generate a pricing model for wool kimbal curtis and john stanton

Using NeuralToolsto generate

a pricing model for wool

Kimbal Curtis and John Stanton

Australian Wool Industry

70% of world trade in apparel wool is Australian wool

Unlike other commodities• Each farm lot is fully measured

• Each farm lot has an individual price About 450,000 farm lots sold each

year in Australia Raw wool value of AUD3 billion

annually

Wool prices & market reporting

Estimates of auction price on individual lots needed by sellers (farmers)

Forecast auction price on individual lots required by buyers for contracts

Market reporting of price paid for different wool types

Neural nets & wool prices

Neural nets attractive because• Number of records is large

• Prices are dynamic• Price/attribute relationships are non-linear with

interactions

• Price/attribute relationships change over time

• The data set is incomplete and imprecise

All Merino fleece lots

(Fremantle Jan-Mar 2006)

Each grey dot represents a parcel of wool sold at auction i.e. a ‘case’

Long & short fleece lots

(Fremantle Jan-Mar 2006)

Long and short wool differentiated on price

Merino pieces lots

(Fremantle Jan-Mar 2006)Pieces wool

(a subset of the wool clip)

Changes to price diameterrelationship (September)

2001 2003

2005 2007

The Challenge !

(Fremantle Jan-Mar 2006)

Market Indicators

Market indicators, like a stock market index, used to price wool

Model development

Stages 1. Assemble 6 month data set2. Use Best Net Search3. Evaluate predictive capability4. Refine model

Model development (1)

Assemble 6 month data setIndependent category and numeric variables

Dependent numeric variable (price)

Training, testing and prediction data

Use Best Net Search Evaluate predictive capability Refine model

Model development (2)

Assemble a 6 month data set Use Best Net Search

GRNN – proved best in most cases(generalised regression neural net)

MLFN – also tried with up to 5 nodes(multi layer feed-forward neural net)

Evaluate predictive capability Refine model

Configuration summary

Net Information Name Net Trained on Pieces wool sales, weeks 33 -

38, 2006 (3) Configurations Included in Search GRNN, MLFN 2 to 3 nodes Best Configuration GRNN Numeric Predictor Location Palisade Conf Curtis v6 BNS 6hrs.xls Independent Category Variables 8 (Sale centre, Sale week, Sale outcome,

Style, Med Hard Cotts, Unscourable Colour, Jowls, Dark Stain)

Independent Numeric Variables 8 (Staple Length, Staple Strength, Vegetable Matter, Diameter, CV Diameter, Mid Breaks, Yield, Hauteur)

Dependent Variable Numeric Var. (Clean price)

Model development (3)

Assemble a 6 month data set Use Best Net Search Evaluate predictive capability Refine model

Model evaluation (1)

NeuralTools outputsError measures

Actual versus Predicted, Residuals

Variable Impact Analysis

Live Prediction Relationships between variables Compare to published market

indicators

Model evaluation (1)Training and Testing summary

Training

Number of Cases 5910 Training Time (h:min:sec) 0:39:43 Number of Trials 104 Reason Stopped Auto-Stopped % Bad Predictions (5% Tolerance) 14.7377% Root Mean Square Error 24.72 Mean Absolute Error 16.42 Std. Deviation of Abs. Error 18.48Testing

Number of Cases 1507 % Bad Predictions (5% Tolerance) 43.3975% Root Mean Square Error 53.18 Mean Absolute Error 36.99 Std. Deviation of Abs. Error 38.21

Model evaluation - Training data(mean absolute error 16 cents)

Model evaluation - Testing data(mean absolute error 37 cents)

Model evaluation (1)Testing data (indicators)

Observed versus predicted for the published Pieces Market indicators

Most points are on the 1:1 line, but a small group hover above i.e. they have higher predicted values than reported

Model evaluation (1)Variable impact analysis

Relative Variable Impacts

41.3%18.7%

11.7%8.8%

7.7%1.9%1.8%1.6%1.2%1.2%1.1%0.9%0.7%0.6%0.4%0.4%

0% 10% 20% 30% 40% 50% 60% 70%

Diameter Vegetable Matter

Staple Length Jowls

Hauteur Sale outcome

Med Hard Cotts Yield

CV Diameter Staple Strength

Sale centre Sale week Dark Stain

Style Unscourable Colour

Mid Breaks

This is a sensitivity analysis, not the percent of varianceaccounted for by each variable

Model evaluation (2)

NeuralTools outputs• Error measures

• Actual versus Predicted, Residuals

• Variable Impact Analysis

Live Prediction Relationships between variables Compare to published market

indicators

Model evaluation (2)Live prediction

Sale centre FremantleSale week W38

Style Average

Med Hard Cotts C0Unscourable Colour H0Jowls J0Dark Stain S0

Diameter 20.0

Yield 50.0Vegetable Matter 2.5

Staple Length 80Staple Strength 35Mid Breaks 55Hauteur 62

Clean price 664

Simple spreadsheet pricing tool.

Change any of the values in the yellow cells, and ‘Live prediction’ updates the clean price

Model evaluation (3)

NeuralTools outputs• Error measures

• Actual versus Predicted, Residuals

• Variable Impact Analysis

Live Prediction Relationships between variables Compare to published market

indicators

Model evaluation (3)relationships between variables

65

70

7580

85

25

30

35

40

45

590

600

610

620

630

640

650

660

670

680

CleanPrice

StapleLength

StapleStrength

SydneyWeek 3821 micron2% VM

Model evaluation (3)relationships between variables

22

mic

ron

21

mic

ron

Fremantle Melbourne Sydney

65

7075

8085

25

30

35

40

45

595

600

605

610

615

620

625

630

635

Clean

Pr ice

Staple

Length

Staple

Str ength

65

7075

8085

25

30

35

40

45

500

520

540

560

580

600

620

Clean

Pr ice

Staple

Length

Staple

Str ength

65

7075

8085

25

30

35

40

45

600

610

620

630

640

650

660

670

680

Clean

Pr ice

Staple

Length

Staple

Str ength65

7075

8085

25

30

35

40

45

590

600

610

620

630

640

650

660

670

680

Clean

Pr ice

Staple

Length

Staple

Str ength

65

7075

8085

25

30

35

40

45

560

570

580

590

600

610

620

630

640

650

Clean

Pr ice

Staple

Length

Staple

Str ength65

7075

8085

25

30

35

40

45

540

550

560

570

580

590

600

610

620

630

640

Clean

Pr ice

Staple

Length

Staple

Str ength

Model evaluation (3)relationships between variables

65

7075

8085

25

30

35

40

45

705

710

715

720

725

730

735

740

745

750

755

760

Clean

Pr ice

Staple

Length

Staple

Str ength65

7075

8085

25

30

35

40

45

750

760

770

780

790

800

810

820

830

840

Clean

Pr ice

Staple

Length

Staple

Str ength65

7075

8085

25

30

35

40

45

740

750

760

770

780

790

800

810

820

Clean

Pr ice

Staple

Length

Staple

Str ength

65

7075

8085

25

30

35

40

45

640

645

650

655

660

665

670

675

Clean

Pr ice

Staple

Length

Staple

Str ength65

7075

8085

25

30

35

40

45

670

680

690

700

710

720

730

740

Clean

Pr ice

Staple

Length

Staple

Str ength65

7075

8085

25

30

35

40

45

660

670

680

690

700

710

720

Clean

Pr ice

Staple

Length

Staple

Str ength

20

mic

ron

19

mic

ron

Fremantle Melbourne Sydney

Price spread variation

Model evaluation (4)

NeuralTools outputs• Error measures

• Actual versus Predicted, Residuals

• Variable Impact Analysis

Live Prediction Relationships between variables Compare to published market

indicators

Model evaluation (4)predictive capability

MelbourneWeek 38

20 micron indicator

22 micron indicator

Model evaluation (4)predictive capability

MelbourneWeek 38

Dark blue lots have SL, SS and VM “similar” to market indicator definition

Model evaluation (4)predictive capability

MelbourneWeek 37

Model evaluation (4)predictive capability

MelbourneWeek 37

Model evaluation (4)predictive capability

MelbourneWeek 36

Model evaluation (4)predictive capability

MelbourneWeek 35

Model evaluation (4)predictive capability

MelbourneWeek 34

Model evaluation (4)predictive capability

MelbourneWeek 33

Model evaluation (4)predictive capability

FremantleWeek 37

Model evaluation (4)predictive capability

FremantleWeek 38

Model development (4)

Assemble a 6 month data set Use Best Net Search Evaluate predictive capability Refine model

• Reduce variables

• Combine selling centres

• Sale week - category variable

Some Neural Net applications

Market reporting Price predictor Validation check for other estimates Missing sale problem Generate price matrices Estimate premiums and discounts

Premium for “organic” wool

800

1000

1200

1400

800 1000 1200 1400

Actual price

Pre

dic

ted

pri

ce

800

1000

1200

1400

800 1000 1200 1400

Actual price

Pre

dict

ed p

rice

June-July saleApril sale

Summary

Data rich application with characteristics that looked ideal for NeuralTools

Solutions generated which can support industry analysis and generation of indicators