Team 4 Amit Tyagi Harmanjit Singh Mohini Jain Nandini Chandrasekhar Rahul Chakraborty

Executive Summary This document attempts to forecast the prices of cement from changes in the price of crude oil. The analysis would help construction companies and other involved players manage their inventories effectively and influence their purchasing decisions. The data for the study, namely oil and cement prices, were collected from syndicated agencies and extend over a 3 year period, from 2009-2011. The methodology included data cleansing and conformity wherein missing values were accounted for by linear extrapolation. The forecasting methods used included Nave, Holt winter exponential and regression (based on multiple predictors like oil and time function) and were compared on accuracy measured by minimal errors in the forecasted values. The data for oil was observed to have an increasing pattern over time while the cement prices had a periodic high every April. The key findings validated this higher price (April) which necessitates inventory replenishment during December when cement prices are comparatively lower. Another important finding was the price of cement was a function of the 1 month lag in prices of oil. The caveats which need to be highlighted are the limitations in data collection, wherein larger samples could have helped us in validating the model better. Ancillary variables like the seasonal rainfall or economic indicators could also have had an impact in forecasting cement prices.

Problem Description The cost structure of the construction industry is dominated by the usage of raw materials such as sand, cement, steel etc. The price of raw materials fluctuates and this fluctuation could be driven by one of the external factors crude oil price. The sudden surge in raw material price adversely impacts the industrys bottom line. The objective of the project is to factor into the change in crude oil price to predict on a quarterly basis the price change in cement, which is one of the key raw materials used in construction. The forecasting model will serve as a valuable tool to builders/dealers in the construction industry or cement production companies to plan capacity expansion.

The benefit of the model will facilitate timing of purchase, inventory management, and budget allocation. Thus, the outcome of the model would include better resource planning and sourcing leading to increased savings.

Data Data Collection Monthly crude oil and cement prices were collected from the website (www.indiastat.com) for the time period January 2009 to May 2011, Data Issues The initial data had the following issues: 1. Missing values for cement for the months January to March 2011 2. Oil prices were available in the US dollar. Issue Resolution Approach: 1. The former issue was solved by imputing missing data using moving average and straight line methods 2. The latter issue concerning with oil price in the US dollar value was converted to Indian Rupees using historic conversion rate corresponding to each data point sourced from www.oanda.com. Data Exploration Observing the data for cement over a 3 year period from 2009 to 2011, we can observe the following from the line chart (figure 3): o The data does not reflect a clear trend. o The annual data shows a peak around April indicating a clear seasonality o Represents a stationary series Observing the data for oil over the corresponding period, we can observe the following(figure 4):
3

o The data displays an increasing trend o No significant seasonality since the line chart shows random spurts through the year o Through 2010, the data displayed a stable pattern o Represents a non-stationary series Nave Forecast The nave forecast uses a 12 month lag to predict values. The cement data shows an annual seasonality from the Jan-09 to Jan-11 values. For evaluating the prediction accuracy, we compared the forecasted and the actual values and examined the forecast errors measures, namely: i) ii) iii) iv) Average Error: -0.3 Mean Absolute Error (MAE): 12.68 Root Mean Square Error (RMSE): 16.04 Mean Absolute Percentage Error (MAPE): -0.43%

Results We can see that the Nave method provides a decent model to forecast cement values since the errors are reasonably low.

Exponential Smoothing Method Given that our series of interest, cement prices, is characterized by lack of trend and 12 month seasonality, we decided to explore the method of exponential smoothing. In particular, we decided on using the Holt-Winter No Trend technique available in XL-Miner which is applicable for data that displays seasonality but no trend. Data preparation As first step, we partitioned the data into training data of 24 months and validation data of 5 months. This was in order to test the level of accuracy attained in using the Holt-Winter No Trend exponential smoothing on our dataset. We used the following default values for smoothing constants: i) Alpha=0.2 ii) Gamma=0.05
4

Forecasts We observed that forecasts made based on this method systematically under-forecast the prices for cement. Having checked the model accuracy on the validation data, we reran the analysis on the entire data to obtain forecasts for the required period Jun 2011 Aug 2011. The forecast values and the corresponding charts are provided in the exhibit as Fig. Exponential Smoothing. Comparing to Nave forecast: While both the Nave forecast and the Exponential smoothing forecast systematically underforecast the cement prices, we can say that exponential smoothing is still better based on the lower RMSE number it produces. Next Step: Given the suspected role of crude oil prices, we decided to explore the regression based methods of forecasting which have been described in the following section.

Regression We conducted a series of regression analyses to explore the significance of crude oil prices in predicting the cement prices. Moreover, in order to account for the seasonality displayed in the data, we used indicator variables through which we explored the option of using both quarterly as well as monthly dummies. In conducting the regression analysis, we used the adjusted R square of the model as an indication of the goodness of fit of the model for the data and hence its strength of prediction. However, while iterating for different models of regressions, we used the error measures such as RMSE and average error to compare across the models to find the better one. Regression Model 1: Cement prices, time index and quarter Dependent variable: cement prices Independent variable: time index, 3 quarter dummies for Q1, Q2 and Q3 (Q4 as base) The adjusted R square for the model is 0.3345 Validation data scoring RMSE = 24.60 and Average Error = 19.02 Regression Model 2: Cement prices, time index and months Dependent variable: cement prices Independent variable: time index, 11 monthly dummies with December month as base The adjusted R square for the model is 0.5364 Validation data scoring RMSE = 14.58 and Average Error = 11.31
5

Regression Model 3: Cement prices, change in crude oil and months Dependent variable: cement prices Independent variable: change in crude oil prices, 11 monthly dummies with December month as base The adjusted R square for the model is 0.5964 Validation data scoring RMSE = 15.16 and Average Error = 8.93 Results: The forecast for one step June 2011 cement prices is Point Estimate: 244.67 95% confidence interval is (230.50,258.9)

Insights and Recommendations Key Insights: The key insights based on the analysis conducted are outlined below: April and May have at least 10% higher cement prices than December Cement Prices are affected by previous months change in Oil prices and not oil prices themselves Crude prices have a positive trend overall last two years. Recommendations: Inventory should be brought in December for next fiscal year rather than April/May Track changes in crude prices on monthly basis to deduce future cement prices

Limitations The key limitations of the analysis are given below: o The analysis has been based on a limited number of data points available. To verify the outcome of the analysis, a more robust analysis with a longer time frame should be ideally used. This is difficult given cement is not publicity traded as a commodity in India o The model can also include an economic variable such as rain forecasts to check if there is a correlation between cement prices and rain in the region which might stall construction activity.
6

Appendix

Figure 1: Initial Data of Crude Oil and Cement Prices

Figure 2: Cement and imputed Crude Oil prices as indicated by the box

Figure 3: Time Series of Cement Prices

Figure 4: Time Series of Oil Prices

Figure 5: Nave Forecasts

Figure 6: Exponential Smoothing Time Plot of Actual Vs Forecast (Training Data)

Validation Period

Time

Actual

Forecast

Time Jan. 2011 Feb. 2011 Mar. 2011 Apr. 2011 May. 2011

Actual 237.67 238.56 238.69 280.00 276.00

Forecast 240.10 237.66 233.27 231.33 237.24

Error -2.43 0.90 5.42 48.67 38.76

LCI 219.09 216.65 212.26 210.33 216.23

UCI 261.11 258.67 254.27 252.34 258.24

Figure 7: Times Series Regression (Model 1)

10

Figure 8: Times Series Regression (Model 2)

11

Figure 8: Times Series Regression (Model 3)

12

Autocorrelation of Residual / Data Set #1 1 0.5 0 -0.5 -1 1 2 3 4 Number of Lags 5 6

13