You are on page 1of 16

UK Gas Consumption

Forecasting Model
Introduction

GAS FORECASTING CHALLENGE:


INTRODUCTION

Contents
Data Cleansing
Running the Regressions
Analysing the Output
Constructing a Model
Testing for Accuracy
Limitations
Opportunities for Further Model Development
Questions
Data Cleansing

DATA CLEANSING
Data Sets Provided: Composite Weather Variable & Gas Demand Data

Can you just assume that the data you receive is useable as is?
DATA CLEANSING CONT. Data Cleansing

Data Errors Noted after flagging where dates


dont match:
1). Several Anomalous Data Readings (Removed) ->

2). Data out of Alignment with Weather Data ->


(Realigned)

3). Data Missing on One Side (Deleted)->

4). Duplicated Data (Removed) >

5). Superfluous Data (out of sample)>


Letting these errors slip through will
skew your model and impact accuracy!
TRANSFORMING THE DATA Transforming the Data

Bringing it back to the Question:

Transforming the Data from


Sum Total of All Areas per Day
the Local to National Level:

Mean Average of the CWV


across all areas per day

These form the Predictor and Criterion Variables


RUNNING REGRESSIONS Running Regressions.

Program Used: SPSS

Regression Criteria:
Descriptive Statistics
Included R Change
Casewise Diagnostics
Removing outliers outside 3
95% Confidence Intervals.
ANALYSING THE OUTPUT Analysing Output

Step 1: Assumptions
1). Normal Distribution

2). Linearity

3). Homoscedasticity
Variance along the
regression line seems
the same

4). Ratio Variables


ANALYSING THE OUTPUT CONT. Analysing Output

Step 2: Model Robustness


1). Descriptive Statistics

2). Pearsons Correlation

3). ANOVA

4). Variance Explained -98.6%


Pretty accurate predictions!
ANALYSING THE OUTPUT CONT. Analysing Output

Step 3: Interpreting the Coefficient


Beta Coefficient Value for the relationship
between CWV and Gas Demand

According to this model, for every 1 point increase in the average UK CWV value,
Gas Demand is expected to fall by roughly 162 million units.
CONSTRUCTING A MODEL
Constructing a Model

Building a Predictive Model in Excel


2). Apply this to
each days
CWV to make
predictions.
1). Work out the
product of the
CWV * + Gas
Demand
Constant.

3). Compare the predicted value against the actual value (see overleaf).
CONSTRUCTING A MODEL CONT. Constructing a Model

Represented in a graph:
In-Sample (2015) Out of Sample (2016)
Predicted Gas Demand vs Actual Demand Predicted Gas Demand vs Actual Demand
3,000,000
3,000,000
Thousands

Thousands
2,750,000
2,750,000

2,500,000
2,500,000

2,250,000
2,250,000

2,000,000
2,000,000

1,750,000 1,750,000

1,500,000 1,500,000
1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th 13th 14th 15th 16th 17th 18th 19th 20th 1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th 13th 14th 15th 16th 17th 18th 19th 20th
Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan

In Sample Period Actual Data (2015) In Sample Period Predicted Data (2015) Out of Sample Predicted Values (2016) Out of Sample Actual Values (2016)

It appears that the predicted


line follows the actual values
quite closely in both in-sample
and out of sample models.
TESTING FOR ACCURACY Accuracy of the Model

Percentage Difference Between Predicted and Actual Values


20%

18%

16%

14%

12%
Average
10%
Out of
8%
Sample
6%
(2016)
4%

2% Average
0% In Sample
1st Jan 2nd Jan 3rd Jan 4th Jan 5th Jan 6th Jan 7th Jan 8th Jan 9th Jan 10th Jan 11th Jan 12th Jan 13th Jan 14th Jan 15th Jan 16th Jan 17th Jan 18th Jan 19th Jan 20th Jan
(2015)
Percentage Difference Between Middle and Actual In Sample Percentage Difference Between Middle and Actual Out of Sample

* % difference for model under-predictions were made positive for the sake of providing an accurate average.

On average, over the out-of-sample


period, the model was 6.57% away
from the actual values for out of
sample, 3.43% for the in-sample
period.
CONSTRUCTING A MODEL CONT. Accuracy of the Model

How Accurate is the Model with more Contemporary Data?


August 2017 Data
Predicted Gas Demand vs Actual Gas Demand
600,000

Thousands
500,000

400,000 Average Percentage


300,000 Difference Between
200,000 Predicted & Actual
100,000
Values:
-
5.01%
1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th 13th 14th 15th 16th 17th 18th 19th 20th
Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug Aug

Out of Sample Predicted Values (Aug 2017) Out of Sample Actual Values (Aug 2017)

* Bad Demand Data Removed from Model


MODEL LIMITATIONS Model Limitations

Data for this model, provided by the National Grid, does not measure all gas demand for the UK.
Some demand may be serviced by providers that fall outside of NG distribution networks.

The data only covers residential and small business consumption. Heavy industry/other consumer
categories are not accounted for in these predictions.

The demand data is provided via non-daily-metered readings. This may lead to inaccuracies, which
may be avoided through the use of smart metering technology/automated meter readers.
OPPORTUNITIES TO IMPROVE Model Improvements

Improving Model Accuracy


Identify other factors that could predict gas demand (level of economic growth, availability
of/demand for competing fuels).

Gather data and incorporate these factors into a multiple regression to increase predictive accuracy.

At points where the accuracy of the predictions deviated significantly, try to identify extraneous
events that may have impacted gas demand.

Improving Applied Value


Incorporate the findings of this model into a bigger model that links expected gas demand with
other factors that affect gas price?
Questions

You might also like