Towards Accurate Demand Forecasting

Demand forecasting is critical in todays ‘Customer is King’ scenario. Businesses are juggling customer satisfaction with on-time delivery; larger and dynamic product portfolios on one side and lean inventory management and multiple supplier relationships on the other.  It is a delicate balance - a sudden burst of demand can lead to stock-outs and complicate resource planning with suppliers, and a slump in sales could lead to excess inventory and capital constraints.  The importance of accurate demand forecasting cannot be emphasized more than ever now, with every business trying to meet their customers ever changing demands, and competing with rivals in shortening delivery times while maintaining margins. Accurate demand planning to face volatile business scenarios is the need of the hour.

Some of the key questions that need to be answered during business planning are:

• What is the expected demand for each of the SKU categories over the next few weeks/months (Demand Forecasting)
• What is the minimal inventory, and the right reorder point for each SKU (Safety Stock and Replenishment Strategy)
• How much inventory of each SKU do we need to maintain at a particular point of time (Inventory Optimization)

To answer all these questions, we first need to be able to understand and forecast accurate customer demand. A number of ways, both qualitative and quantitative have been tried and tested, to forecast demand.

Qualitative v/s Quantitative Forecasting

Most qualitative methods adopt use a top down approach where Management decides on targets, and based on marketing inputs, the forecast is broken down to the smallest unit required. Marketing inputs are primarily extrapolations combined with their current market knowledge and specific marketing plans for the particular time period.

An example is the Delphi method developed by RAND Corporation where a group of suitable experts, typically 5-20 members, are asked to give their forecasts and reasons for the same. Discussions continue till a consensus is reached. This method would then take a median or mode of the experts forecast. Another popular technique used is to conduct Customer surveys to evaluate demand based on customers purchasing patterns and preferences.

Most quantitative techniques and statistical models use a bottom up approach based on historical sales by SKUs and other independent metrics that impact sales to forecast demand. While the choice of forecasting method mainly depends on availability of data, statistical techniques are known to give far accurate results.

One of the main statistical methods used for forecasting is the Time Series technique. This method is extrapolative in nature, i.e. it uses trends in historical sales as an input. Time period based factors like seasonality trends that influence sales are also taken into account to help extrapolate sales.

Some of the Time Series methods used are:

- Moving Averages : In this method the forecast is calculated as a moving average of actual demands of ‘n’ number of past time periods. The number of periods to be chosen is a decision to be made based on data. If we have stable underlying pattern then a large value of ‘n’ is acceptable, but if there are short term fluctuations to be seen then a small value makes more sense. Weighted moving averages are also used where each time period is given a weight.

- Exponential Smoothening methods like Holt Winters method : This is a variation of the Moving Averages method. It incorporates time parameters like trend, growth, seasonality of historical data, and gives recent data more weightage. We have single exponential smoothening; double exponential smoothening (which is applied when the data shows a trend), and triple smoothening or the Holt Winters method. The Holt Winters Method is applied when the data exhibits both trend and seasonality.

- Another popular method is ARIMA : Autoregressive Integrated Moving averages method is another application of the moving average models.  Subsequent outputs reflect autocorrelations from both historical data and previous outputs. This method predicts values as a linear combination of the data’s own past values and past errors. ARIMA models can also include other time series data as inputs and these models are called ARIMAX models. These models give best results when data is stable.

Other than time series, Explanatory methods are also employed for forecasting; these methods predict sales (dependent variable) based on other independent variables.  Examples of such models are regression models, Artificial Neural Networks (allows for non-linear relationships between the dependent and independent variables), predictive modeling and other econometric modeling techniques.  These methods are used when sales history can no longer be considered as the sole predictor for future demand.

Recently other set of forecast model that use both time series and explanatory variables are being used these are known as dynamic regression models, longitudinal models or transfer function models.

Building a Forecasting Model

A number of factors need to be taken into account before building a model:

• Interaction with key stakeholders – to answer questions such as what needs to be forecasted, time period for which forecast is required, current methodology for determining forecasts, etc
• Collection of data – Available historical sales data needs to be collected in addition to other data such as sales targets for next month/year, marketing budgets and allocations, marketing plans, etc.
• Available tools – Tools that are currently available in your organization, can be MSExcel, SAS, R  or any other statistical tool.
• Choosing the right model – A starting point can be to plot the available data to determine trends, presence of seasonality, outliers and correlation between available variables. The final model used depends on data availability, strong correlation between explanatory variables with sales, etc. It is common to use more than one model - the committee approach.
• Evaluating a forecast model – evaluation of a forecast model is possible only after actual data is available for that time period, common metrics used are MAE (mean absolute error), MSE (mean squared error) and MAPE (mean average percentage error).

A recent forecasting exercise is shown below.  In this case, the business objective was to forecast daily demand for the next 60 days, for a highy seasonal product.  Forecast accuracy achieved was close to 92%.

A more detailed case study can be found here.

Other Forecasting Tools

There are a number of software packages available in the market that aid forecasting sales like SAP APO, Oracle Demantra, Forecast Pro etc. Each of these packages have their unique way of using the above techniques, treating outliers, and understanding seasonal patterns and trends.

However, it is important to understand how these packages work – what technique is being used, how outliers being treated, what corrections are being made to the data, a business check on the assumptions and how does it impacts the final forecast. Without this basic understanding some packages may seem like a black box that churn out numbers for the business to use.

This leads us to believe that Statistical models that are tailor made and build from scratch, along with business sense and other situation-specific business inputs may be more are reliable and holistic.

How Nestle's former executive Carlo Donati is grooming data analytics startup Marketelligent

Reproduced from The Economic Times article, Oct 16 2012.

Please click here for article.

Having spent 34 years in Nestle, including six years as the head of its South Asia business, Carlo Donati knows a thing or two about who the consumer is and how a company can reach him. The former Nestle lifer also feels companies today are not pushing the right buttons to take decisions. "Data analytics is the missing piece in the top management dashboards," he says. "They can't rely on gut feel to take decisions."

For the 66-year-old Swiss, that intellectual gap is also a business opportunity, one he is tapping by being a mentor, client-builder and, in time, an investor in a $2 million Bangalore-based data analytics startup called Marketelligent. "Data analytics gives the Key Performance Indicators (KPIs) and Marketelligent has identified this niche," says Donati in a telephonic interview from Lugano in Switzerland.

It helps that when he was in Nestle Asia, Roy Cherian, the CEO and one of the two founders of Marketelligent, worked under him for three years. The 47-year-old Cherian headed the chocolates and confectionery business of Nestle India.

While the former Nestle duo knows much about consumer businesses and decision-making, what they don't know as well is the technology that can swallow a mountain of data and spit out pithy observations. And that's where Anunay Gupta, the other founder, comes in.

Gupta heads analytics of Marketelligent, and he comes with a raft of experience in this area. The 44-year-old set up Citibank's analytics back office in Bangalore to support the US bank's cards division. And the Citi global manager who forged the bank's analytics drive, Marcia Tal, will be advising Marketelligent on clients and sectors in which to offer services, as well as technical and business aspects of data management.

Tal, 53, spent 25 years at Citi. During this time, she started and headed Citibank's worldwide analytics division before forming her own consultancy, Tal Solutions. Gupta worked under Tal for five years. "Anunay was one of the first employees of the Bangalore Citi back office. In every sense, it was a startup - inside of a large corporation ( Citigroup)," she says. "The mindset, skills, commitment, tenacity and motivation I saw in Anunay are why I believe he will be successful in building his own company." Tal also advises Black Oak Partners, a US-based data analytics firm.

Marketelligent, formed in 2008, is positioning itself as a data analytics company servicing the retail, banking, energy and healthcare verticals. It will delve into aspects like consumption patterns, market trends and consumer behaviour, among other things. "The founders have competence and professionalism on the topics, which are fundamental for success," says Donati.

According to Cherian, data analytic techniques are transferable across sectors. "You need experts in statistics and forecasting." At the same time, they are looking to hire subject experts, for which they plan to seek help from their two global advisors as well. "We are in talks with an energy expert based in Texas," says Cherian. "He is my vintage (same age), an Indian, and we reached out to him from our IIM network."

Between the founders, Gupta is the techie, responsible for technology choices and delivery quality, while Cherian focuses on operations, marketing and funding. Says Cherian: "The duo - Donati and Tal - help us reach out to global customers." He adds that, since Donati came on board, the company has begun pursuing two client leads in Europe, but he declines to share their names.

While decorated executives as advisors to startups are common, a high degree of engagement in operations is not. Donati is prospecting for clients for Marketelligent - a rare thing for an executive of his stature to do. "Often a marquee customer can help bag business more than an advisor. However, if the advisor prospects for business, it's a different case," says Raman Roy, who has himself birthed four BPO startups.

Sanjeev Aggarwal, MD of Helion Ventures, a venture capital firm, feels the value a person like Donati adds to a startup depends on the level of engagement. "An eminent global name can help in customer acquisition in Fortune 500 companies, which is extremely tough for a startup," he says. "A big name will help open doors, get audience quickly, but a deal is an entirely different matter."