The Heuristic Business Forecasting Methods Revinda and Metrix


Scientific Study, 2015

12 Pages


Excerpt

Table of contents

1 Introduction

2 Brief Forecast-Methodology Overview

3 Business Forecasting

4 Comments on Forecast Accuracy

5 New Methods supporting Business Forecasting
5.1 Business Process Forecasting - The REVINDA-MM-Approach
5.2 "ERP-Type"- Forecasting - METRIX Approach

6 Conclusions:

7 Literature:

Appendix

Summary / Intention of the Paper

The report describes two new heuristic approaches to time series analysis and forecasting for business purposes. Both approaches avoid any assumptions according to assumed process attributes behind the data (e.g. stochastic process, stationarity, normal distribution of ran­dom noise, …). Those methods just engineer data of any kind of business processes. – Only unidentified (inherent) process structures are used for forecasting.

Key words: Revinda, Metrix, Business Forecasting, Similarity, Model-free forecasting Methods

1 Introduction

Speed represents the driver of current business development. IT is about allowing for auto­ma­tic self-synchronizing (production) processes. Big Data potentially offers the identifica­tion of hidden structures enabling for further process improvements. In many companies mobile information access is being used. Multi-Channel B2C, B2B and M2M are gaining the manage­rial pole position. - But nevertheless the qua­lity of data is the key for producing excellent results. We cannot deny that GIGO deter­mines the outcome. So, it is important that planning is based on as 'realistic' data as possible. After roughly more than 35 years Business Forecas­ting is back in the focus.

According to 2 "Industrie 4.0" represents the successor of the CIM-Conception of the late 1970-ties, now based on advanced IT infrastructures. In histo­ry of Business Trends the pre­decessor of CIM was the intro­duction of "quantitati­ve data" for impro­ving business process­ses. At that time many extra­polating fore­casting methods have been develop­ed (Holt/­Win­ter, Lewandowski, etc…), which up to now are standard in estimating future data of business processes. The Helix 1 and Helix 2 cycles 2 describe the deve­lop­ment of business trends. The Helix findings can be interpreted as a structural forecasting approach.

2 Brief Forecast-Methodology Overview

Forecasting represents a challenging task. Each problem is linked to special requirements, restrictions, available data and the expertise of selecting the most appropriate method. Common to all is the need for "good results". There is a great number of forecasting methods available. Some of them are standard for certain categories of forecasting tasks. E.g. most ERP-Systems offer 10 or more (simple) standard forecasting methods for routine planning operations. (In this report this kind of forecasting will be called - in line with present penetra­tion of ERP's - "ERP-Type" forecasting.)

The list below specifies a number of well known forecasting methods:

- Simple Extrapolations (e.g. EXCEL-Functions for Trends)
- Decompositions of Time Series (Level, Trend, Season, Random Noise)
- Smoothing Methods <Filters> like Moving Average (MA), Exponential Smoothing, Holt / Winter extension for Trend and Seasonality
- Box-Jenkins Method and advanced auto-correlated ARMA and ARIMA-Models and ARCH, GARCH
- All kinds of Regression Analyses (Least Square Estimators)
- Econometric Modelling (Kalman-Filter, Markoff-Chains, …)
- Neuronal Networks (MLP, SOM, …)
- Croston-Method for intermittent/sporadic demand
- Diffusion Models (Bass-Models) for lifetime analyses
- Black-Scholes Equation for financial analyses
- Evolutionary Approaches (Delphi-Method, Scenario Planning)

The list – of course – is not complete. Different software packages (e.g. "R", "MATLAB", …) offer sets of standard forecasting approaches and /-models.

3 Business Forecasting

This paper covers forecasting approaches dedicated for improving planning quality. As plan­ning is the backbone of operational business success, forecast quality is put in focus. Roughly spoken there are 2 main categories of forecast scenarios now being discussed:

In this paper Business Forecasting distinguishes 2 sub sets:

a) Business Process Forecasting (MM-Forecast) - see 1
b) "ERP-Type" - Forecasting

a) Forecast as a Business Process (MM-Forecast1)

To give a B2C example: A LSP2requires from his customer (a B2C company) the forecast of daily picks for two weeks ahead for capaci­ty planning, because the 24 hours service only is guaranteed in case the forecast meets the range between 75% and 120% around actual orders. – Of course the B2C-company knows the order pattern from past experience. But daily B2C orders are influenced by standard and special promotions of Marketing. Another aspect affecting daily sales is given by temporary top selling items and/or the launch of new products, etc….The effects of pro­motions and special market conditions are estimated 'manually' by Marke­ting.

So, the final forecast has 2 input sources: a "basic process forecast" and the impact of special com­pa­ny activities. Under those scenario conditions a manager will be responsible for the final forecast quality and data maintenance. So, the final fore­cast summarizes the basic forecast plus – in this example – Marketing "brain-input". – What about the data to be used in next year's forecasts? As top selling items and special pro­motion will not occur next year at the same period the actual data have to be re-adjusted for providing high quality input data for next year's basic process forecasts. Therefore it makes sense documenting the promotional and special market effects in a log-file. In case of daily forecasts the re-adjustment of weekly data is sufficient.

The same organization holds in case the company also runs a retail chain – e.g. with some hundred outlets – to be delivered on a daily basis as well (B2B). In this case additional factors also will have to be conside­red, e.g. the volume of allocations fixed by the central purchasing department which again represent part of the "brain-input" to the final forecast.

b) "ERP-Type"-Forecasting

The second scenario represents the procedure of applying standard forecast methods being provided e.g. by ERP-Systems. Standard ERP-Systems offer a list of 10 or more diffe­rent standard forecasting methods. During the first implementation of the ERP-System the most appropriate forecasting method(s) are carefully selected for going live. But those forecasting methods mostly remain active even the business conditions have changed after some time. The complexity of ERP-systems often represents the key obstacle for re-adjusting the selection of "best" forecasting methods in time. So, the quality of forecast data is suffering and thus the complete planning process.

The above scenarios usually are applied on calendar-periodical processes, analyzing time se­ries according to decomposition approaches for Level, Trend, Seasonality and Random noise. (Some ERP-Systems also offer forecasting methods [like the Croston-Method] for intermit­tent / sporadic demand).

4 Comments on Forecast Accuracy

This paper is dealing with short term forecast accuracy only, i.e. 1-3 periods (months/weeks) ahead. Daily forecasts are calculated as break-downs from weekly forecasts – based on daily patterns which might change over time within the calendar periodicity.

The standard measures of forecast errors (MSE, MAD, MAPE and TS) will not be discussed here. For the methods we are going to present in this paper MAPE (ti) will be used and the Tracking Signal (TS) will be compared with a new "Structural Tracking Band" based on historical data).

All forecast approaches to be presented in this paper are based on data attributes and pat­terns iden­tified from existing historical data only. No modeling is applied. For identifying dif­ferent cate­­go­ries of attributes a special tool-kit has been developed including relevant statis­tical tests. Let's take two extreme examples: (1) Historical data represents a (periodic) func­tion; e.g. f(x) = A*sin(a*x +b). This implies that there is no random noise at all [ε= 0]. Do we expect that the forecast approach can identify the structure reacting with forecasting the same function (MAPE(tn) = 0)? In the other example (2) historical data represent a ran­dom process [ε = 100%] with for instance Xt ϵ ℝ[A; B] ∀ t - or any other distribution. What about the structure of the resulting forecast data and MAPE(tn)? Answers will be given as far as available when dis­cus­sing the methods in detail according to ongoing research.

The problem behind is the question: Is it possible – and IF, how – to forecast forecasting ac­cu­racy for Business Processes, givenin-samplehistorical data sets without specified (sto­chas­tic or other) pre-condi­tions? From examples it can be seen, that correlation – especially in calendar-periodicalin-sampledata enabling comparison of two past data periods – represents a rough determiner for complexity and predictability. J. Garland, R. James and E. Bradley 3 present an overview about model-free quantification of time series predictability based on entropy. They show that predictability might not be fully exploited by the forecasting method applied from modeled relation between WPE (Weighted Permutation Entropy) and MASE (Mean Absolute Scaled Error). For more details using entropy as measure of complexity of time series see 4, 5 and 6.

Periodical surveys (e.g. M-1/-2/-3-Competitions in 1982, 1993 and 2002) - see 7, 8, 9, 10 – compare the accuracy of forecasting methods being applied on a variety of data sets. Those surveys highlight specific forecasting problems and (most) appropriate model/method selec­tion. Interesting comments on the results of M-Competitions related to business forecasting are given by S. Kolossa in 11.

The purpose of this paper is contributing to solutions of practical (business) forecasting problems by providing simple heuristic approaches (aside of scientific approaches) following the practitioners' principle: "Nearly right beats exactly wrong!"

McCarthy et al. 12 show that in spite of increasing computational power forecast accuracy seems to be deteriorating. Unfortunately no comparison between 2006 and now has been found. The reasons for deterioration seem to be globalization of markets, increasing complexity and speed of business processes, de­crea­sing customer-/ and brand loyalty, shorter product life cycles and others more. On the other hand other surveys show that for selecting forecast tools forecast accuracy remained the priority 1 as well for managers as for practitioners 11 over the last 2 decades.

5 New Methods supporting Business Forecasting

5.1 Business Process Forecasting - The REVINDA-MM-Approach

The name of the approach represents a short cut of "ReverseIndexData Transcription".

Forecast options:

REVINDA has been designed for short-term forecasts, i.e. up to 3 periods ahead. This means 3 months or 3 weeks ahead. The weekly forecast can be fractionized to daily forecasts using daily patterns. Calculation of daily patterns will not be presented here due to complexity of calendar synchronization (moving holidays, etc…).

In addition, the design of the REVINDA approach allows for providing the set of fore­cast values for the complete forecast-period H0 on top of 1 -3 periods ahead.

Data requirements:

Two basic periods of calendar-periodic data are required. Business forecasting allows to restrict the periodicity to subsets of data (e.g. monthly data for 2 years, weekly or daily data ac­cor­dingly). In case of inten­ded daily forecasts, the year will be represented by 13*4 weeks for ca­lendar synchronization rea­sons.

Notations:

1 n, t describe the basic period length (e.g. n=12 or 13 for years; n=52 for weeks; and 365/ 366 for days), while {t=1, 2, …, n} represents the time axis.
2 Historical periods are denoted by H_(-2) and H_(-2) while H_0=H represents the calendar forecast period.
3 Historical Data are denoted by {x_(-2,t) ϵ H_(-2) and x_(-1,t)ϵ H_(-1) ; t=1, 2,…, n}; Forecast Values by {x ̂_t∈ H_0 ; t=1, …,k}..

Algorithm:

1 Cleaning original (raw) data from outliers represents an admissible option.
2 Transformation (1) of historical data to increments of (suitable) reference functions3 f_(-2) (t) and f_(-1)(t) related to historical data. Trans¬formed data are labeled as P-Index; P_(-2,t) and P_(-1,t) ; e.g. P_(-2,t)=x_(-2,t)/f_(-2) (t).
3 Transformation (2) of historical data to sequential increments named S-Indices; S_(-2,t) and S_(-1,t) ; e.g. S_(-1,t)=x_(-1,t+1 )/ x_(-1,t). For a 1 period forecast S_(-1,t)=x_(-1,t+1 )/ x_(-1,t) and S_(-2,t)=x_(-2,t+1 )/ x_(-1,t) ;∀t are calculated. For a 2 period forecast S_(-1,t)=x_(-1,t+2 )/ x_(-1,t) and S_(-2,t)=x_(-2,t+2 )/ x_(-2,t) ; etc. … In general: S_(-2,t+j)=x_(-2,t+j+1 )/ x_(-2,t) and 〖 S〗_(-1,t+j)=x_(-1,t+j+1 )/ x_(-1,t) for {j=1,2,3};
4 P_(0,t)=P_t = a1*P_(-2,t) + b1*P_(-1,t) ; a1, b1 representing Revinda-Coefficients for minimizing forecast errors.
5 Analog step 4 the weighted averages for S_(0,t+j)=S_(t+j) for {j=1, 2, 3} are calculated.
6 P ̂_(t )=〖P_(t )*f_0〗_ (1) ; S ̂_(t+j)=P ̂_(t )*S_(t+j) {j=1, 2 , 3}.
7 Final Forecasts x ̂_(t+1 )represent a linear combination of P ̂_(t+1) and S ̂_(t+j) for {j=1, 2, 3}.

Parameter specification

Specification of the parameters involved seems to a problem. But that's not true. There is a standard set of initial parameters to be applied for all time series. Those parameters result from experience with many time series. Before using data for forecasting with REVINDA the historical data are analy­zed – using a self-made system-tool – including relevant statistical tests. The tool identifies main attri­­buts of the historical time series.

Starting the forecasting process the first forecasts are calculated using the standard parameter values. – In case the forecast accuracy does not meet accuracy targets the parameters will be adjusted according to "optimization criteria". The user is free selecting an accuracy target function. E.g. "Minimize the MAD along the past k forecasting periods". Any other target function is feasible. This procedure represents a WHAT-IF kind of simulation. Based on Newton's Iteration or Gradient-Iteration local minima will be identified resulting in a new set of parameters –comparable to the SOLVER-Function in EXCEL. Experience has shown that target functions inclu­ding the P_0-Index, which is available for the complete H_0 period, show best results. – Thinking 'down the road' this approach represents a cybernetic cycle adapting the historical (data) process to meeting actual values according to the selected optimization criteria. But the resulting 'new' historical process is not unique. Many solutions are valid. Finding the criteria for selecting the "best" new historical process might challenge research.

Practical experiences with REVINDA

REVINDA enjoys about 3 years application for a rather big retail chain running about 300 outlets in Germany for daily forecasts 2 weeks ahead. As a service for the LSP daily B2C orders and B2B orders from the outlets are fore­casted for 24 hours delivery service – given accuracy targets on a daily basis of 75% - 120% of Actuals. The company has established the organization of Business Process Forecast as des­cribed in Section 5a. REVINDA is being applied on a weekly basis. - Adjustments of the parame­ters were needed just twice within 3 years. A visual display is given in Appendix 1 (3).

[...]


1Man-Machine Forecast

2LogisticsServiceProvider

3 For all periods H_(-2) and H_(-1) and H_0 the selected function-type of f(t) (e.g. linear, polynomial, log, exponential, …) remains the same. The shift in y-direction for each period depends on the time series data. For some categories of business time series (e.g. Demand, Orders, Sales, …) f_0 (t) can be conjoined with company plans.

Excerpt out of 12 pages

Details

Title
The Heuristic Business Forecasting Methods Revinda and Metrix
Author
Year
2015
Pages
12
Catalog Number
V542856
ISBN (eBook)
9783346155313
ISBN (Book)
9783346155320
Language
English
Tags
business, forecasting, heuristic, methods, metrix, revinda
Quote paper
Klaus Spicher (Author), 2015, The Heuristic Business Forecasting Methods Revinda and Metrix, Munich, GRIN Verlag, https://www.grin.com/document/542856

Comments

  • No comments yet.
Read the ebook
Title: The Heuristic Business Forecasting Methods Revinda and Metrix



Upload papers

Your term paper / thesis:

- Publication as eBook and book
- High royalties for the sales
- Completely free - with ISBN
- It only takes five minutes
- Every paper finds readers

Publish now - it's free