Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Forecasting Analytics
Topic: Forecasting Fuel Consumption At Nation Level – US
Group Project Report (Proposed By):
Team B8 ; Section: FCAS-B
Name PG ID DHRUV GOEL 6171-0938
NAIREET GHOSH 6171-0078
DIWAKAR SANDUJA 6171-0727
PRABHJOT KAUR 6171-0710
AMEY RAJPUT 6171-0570
SRI RAM 6171-0592
Executive Summary
Business Problem Description: To help Valero Energy Corporation (a fortune 500 company
involved in manufacturing and marketing of fuel and petrochemical products / power) to
effectively plan for production planning, strategic procurement, pricing contracts and
financial hedging and investment strategies for 2016.
Forecasting Objective: To design a forecasting model for accurately predicting consumption
of different refinery products in the US for 2016. The intent is to use consumption patterns
to plan & optimize operations, foresee cash flows and manage procurement of inventory .
Brief description of data : The data contains 22 years of historical monthly data on different
types of US refinery products including price information
Source: US Energy Information Administration (Forms EIA-782C, “Monthly Report of
Prime supplier sales of Petroleum Products for local consumption”
Key Characteristics: Month – Year Level ; Available Period: Jan 1983 to Dec 2015
Series to be analyzed (Observation period 2001 to 2015)
o Total, Regular and Premium Gasoline
o Aviation Gasoline, Jet fuel and Fuel oil
Components observed: Level, Seasonality and Trend (With Noise)
High level description of the final forecasting method and performance on meaningful
metrics (compared to benchmark)
Short Term Horizon: The final model is based on running the ARIMA model over Multiple
Logistic Regression (MLR) by removing auto-correlation providing us forecasts for the 1st
two months in 2016 with high accuracy.
Long Term Horizon: The final model is based on running the Holt’s winter which provides us
with forecasts of 12 months of fuel consumption with a relatively low accuracy or MAPE
Conclusions and Recommendations
The short term model should be re-run every month for prediction of next month’s
consumption.
The long term model should be re-run every 12 months for prediction of next 12
month’s consumption.
External factors such as environmental, political, and economical risks are not
accounted in the model
Price and profitability should be considered in conjunction with the consumption
pattern and prediction to plan production accordingly
Technical Summary
Data Preparation / Issues As quality of data is critical for forecasting results, it is important to
refine and prepare data to make it usable for analysis. The following aspects of collected data
were checked:
Data sources – multiple or single/future sustainability/frequency/ Extreme Values
We have used a single data source. Hence there are no expected issues pertaining to
inter-mixing of different data sources. As this data is periodically shared by U.S.
government, sustainability of data shall not be a problem. Monthly frequency is the
optimum granularity for this analysis as the monthly aggregation averages out
fluctuations. Also a monthly forecast is most relevant for action-ability and agility.
Missing values and Choice of time span
We collected monthly data from Jan-1983 to Dec-2015. We encountered missing values
in the 1980s and 1990s. Therefore we limited our time span to 2001 to 2015 which does
not have the issue of missing values.
Forecasting Methods with details
As the preliminary investigation by chart-plotting revealed the presence of Level, Trend and Seasonality, therefore we have considered forecasting methods that consider both trend and seasonality. The following predictive models were performed and their performance evaluated to arrive at the final model
Forecasting Methods and Performance Comparison to benchmarks – Aviation Fuel (Charts for
other fuel types added in Appendix):
Long Term (Holts’ Winter Model): To obtain forecasts of next 12 months (MAPE: 11.7%)
Short Term: To predict accurate results for in the short term, we ran MLR followed by ARIMA
(3) which led to decrease in MAPE and more accurate results for latest 2 months.
Forecast
Residuals show no trend or
seasonality left over after Holts’
Winter Model
Forecasted only 2 months of data
in 2016 with improved MAPE of
11.3
Residuals show no trend or
seasonality left over after Holts’
Winter Model
Appendix ( Results for the Other Series)
Jet Fuel (Long Term: Holt Winters’ Model)
Jet Fuel (Short Term: MLR + ARIMA (3)):
Residuals show no trend or seasonality left over after
Holts’ Winter Model
MAPE : 2.3
Residuals show no trend or seasonality left over after
ARIMA(3)
MAPE : 2.3
Forecast generated for only 2 months using ARIMA (3)
Fuel Oil (Long Term: Holt Winters’ Model)
Fuel Oil (Short Term: MLR + ARIMA (3)):
Forecast generated for only 2 months using ARIMA (3)
Residuals show no trend or seasonality left ARIMA(3)
MAPE : 19.3
Residuals show no trend or seasonality left after Holts
Winter
MAPE : 21.9
Total Gasoline (Long Term: Holt Winter’s Model):
Total Gasoline (Short Term: MLR + ARIMA (3)):
Residuals show no trend or seasonality left over after
Holts’ Winter Model
MAPE : 3.3
Forecast generated for only 2 months using ARIMA (3)
Residuals show no trend or seasonality left ARIMA(3)
MAPE : 1.8
Regular Gasoline (Long Term- Holt Winter’s Model)
Regular Gasoline (Short Term):
Residuals show no trend or seasonality left over after
Holts’ Winter Model
MAPE : 2.6
Forecast generated for only 2 months using ARIMA (3)
Residuals show no trend or seasonality left ARIMA(3)
MAPE : 1.6
Premium Gasoline (Long Term – Holt Winter’s Model)
Premium Gasoline (Short Term- MLR + ARIMA (3) Model)
Residuals show no trend left over after Holts’ Winter Model
MAPE : 11.1
Forecast generated for only 2 months using ARIMA (3)
Residuals show no trend or seasonality left ARIMA(3)
MAPE : 2.2
ACF Plots