1. OEM Parts Demand ForecastingSwagat Ranjan Behera
Contents
• About data, and other observations
• Business objective
• Box plot for quick understanding
• Time series (Decomposition) plots
• Models
• Discussion of results, and
• Accuracy
OEM Parts Demand Forecasting
2. 2OEM Parts Demand ForecastingSwagat Ranjan Behera
Time Series Period - Jan-06 to Sep-12
Variable Name Usage in Model
(Time Frequency) Year & Month
OEM Parts One Sales
OEM Parts Two Sales
OEM Data – Description, Business Objective
Business Objective:
Forecasting of OEM components demand by using historical sales.
To identify a better model which will forecast the future demand based on the sales data in order
to be well prepared for supply as per demand needs.
Explicitly, specify, forecast accuracy for each modelling technique and reasons for its adaption.
Hence, study carried out with given data provided above and apply best model that can capture all
seasonal and trend components for the provided data.
The data is all about automotive domain
Original Equipment Manufacturing (OEM)
components sales.
Provided sales of two components and
understood them as parts of high-end SUV.
Hence, there understanding is must for further
steps.
3. 3OEM Parts Demand ForecastingSwagat Ranjan Behera
OEM Components Sales Data & Its Understanding
Box-Plots:
Fig.1: Box Plot - OEM Sales Component one
Beside plot (Fig.1) exhibit summary of sales of
component one, where in it shows for few months
(3 in number) there exists zeros, i.e. no sales
happened for this component.
Also, one can observe few outliers, though they
are not very extreme and may represent seasonal
variation, same can be confirmed through
seasonal plots.
Coming to (Fig.2), component two sales exhibit
positively skewed distribution in nature having
median almost near to upper quartile.
Further, for this component sales, we see less
variation compared two first one, where most of
the values are lying between 30+ to 40+.
Fig.2: Box Plot - OEM Sales Component Two
4. 4OEM Parts Demand ForecastingSwagat Ranjan Behera
Time Series Decomposition Plot of Component One
Decomposition Plot (Original Series):
The sessional plot says that there is spike in sales in the month of January, and higher
seasonal variation as per grey bar in the right side of its box compared to trend.
In the month July 2008 the sales touched bottom (trend plot) and immediately
increased from there till Feb 2009 (reached maximum sales) and then the sales
started decreasing. Again the sales started picking up in July 2012, resembling some
cyclic trend.
5. Time Series Decomposition Plot of Component Two
Decomposition Plot (Original Series)
The component two sales sessional plot says that there is spike in sales in the
month of August every year.
Further, trend plot shows that in the month Feb 2009 the sales came touched
bottom and there after sales started to increasing.
6. Auto Correlation Function (ACF) Plots
ACF Plot (Component One) – Fig.3
The component one sales ACF plot
(fig.3) says though there is some
auto-correlation it is not strong to
difference the series.
However, we tried to difference
the series and check whether it is
making sense, but, differencing
made it worse (over differencing)
and couldn’t help much.
ACF Plot (Component Two) – Fig.4
The component two sales ACF plot
(fig.4) clearly exhibits that series is
auto-correlated and stationary,
hence it needs to be differenced.
Hence, differenced the series
before fitting models.
7. 7OEM Parts Demand ForecastingSwagat Ranjan Behera
Forecasting Model Approach
Frequency
(Monthly) Historical
Values as Input
System RAM
Capability (3 GB)
Models &
Functions
Results & Accuracy
Framework followed for developing ‘Demand Forecasting’.
1 Component One
2 Component Two
• ARIMA
• Exponential
Smoothing
• DL (LSTM)
8. 8OEM Parts Demand ForecastingSwagat Ranjan Behera
Discussion of Results – Component One (slide 1 of 2)
Demand Forecasting Results:
Month & Year Actuals Arima ETS LSTM (DL)
Sep-12 9 10.43 7.66 7.65
Oct-12 7 12.35 7.66 7.65
Nov-12 19 13.88 7.66 5.06
Dec-12 18 15.08 7.66 20.47
Jan-13 21 16.04 7.66 19.98
Feb-13 19 16.79 7.66 18.65
Mar-13 16 17.39 7.66 10.99
Apr-13 12 17.86 7.66 16.66
May-13 13 18.24 7.66 14.07
Jun-13 22 18.53 7.66 20.67
Jul-13 23 18.77 7.66 22.75
ARIMA 3.84
ETS 8.73
LSTM (DL) 2.92
Mean Absolute Error (MAE)
Forecasting plot (Actuals vs. Predicted):
We have presented, ARIMA, ETS, and LSTM (DL) forecasts results above for both discussion of their
accuracy and fitment to the business requirement.
On the right hand side, I have presented all forecasting figures and mean absolute error (MAE)
with respect to actuals, and their MAE’s with respect to each model. And, one can observe that
MAE being 2.92 is lowest provided by LSTM (DL) model.
That is, ARIMA has provided curvy linear forecasts which are closer to actuals but didn’t exhibit
same patterns as actuals, however, LSTM (DL) provides more accurate forecasts though over fitted
but with lesser MAE.
9. 9OEM Parts Demand ForecastingSwagat Ranjan Behera
Discussion of Results – Component Two (slide 1 of 2)
Demand Forecasting Results:
Month & Year Actuals ETS LSTM (DL)
Oct-12 43 42.86 40.33
Nov-12 45 42.86 40.33
Dec-12 45 44.86 39.63
Jan-13 46 44.86 39.39
Feb-13 45 45.86 39.39
Mar-13 45 44.86 39.45
Apr-13 44 44.86 39.53
May-13 44 43.86 39.45
Jun-13 46 43.86 39.53
Jul-13 46 45.86 39.39
ETS 0.78
LSTM (DL) 5.26
Mean Absolute Error (MAE)
Forecasting plot (Actuals vs. Predicted):
We have presented above, ETS, and LSTM (DL) forecasts results, however, ARIMA couldn’t produce
any forecasts for this series, hence, we couldn’t present its results.
On the right hand side, I have presented all forecasting figures and mean absolute error (MAE)
with respect to actuals, and their MAE’s with respect to each model. And, one can observe that
MAE being 0.78 is lowest provided by ETS model.
That is, ETS has provided more nearer forecasts than DL (LSTM) as this series is not having more
linear patter and revolving around an average sales of 45 for the test data set.
10. 10OEM Parts Demand ForecastingSwagat Ranjan Behera
Summary
Here, the objective is to forecast the OEM (Original Equipment Manufacturing) component
sales demand based on their historical data.
And, we got two component sales, one is with high seasonal variation and other has flat
sales.
From, ACF plot it is observed that the component one doesn't need any differencing and
component two need differencing.
We have developed all together 3 models: Linear Approximation Models like 1. ETS, 2.
ARIMA and 3. LSTM (Long Short Term Memory) using deep learning framework (non
Linear).
For component one forecast we observed that both linear approximation model (ETS and
Arima ) either provided flat forecast or a curvilinear forecast. However, LSTM (DL method)
provided better forecasts though it seems to be overfitted.
For component two ARIMA even couldn't produce any appropriate results where as ETS has
produced flat forecast, however, this time LSTM (DL) has provided under forecast results.