Intelligent Demand Planning Using AI/ML

Evolution is relentless; change is intrinsic to life and business. It is important to note that some things need to adapt quickly while others can afford to take some time. In this context, we can be certain that Supply Chains fall in the first category. They need to be highly adaptive and resilient to shock. However, a supply chain can be considered as adaptive only if all the elements within the chain are adaptive – these include demand planning, supply planning, inventory planning, production planning, etc. Supply chains are most vulnerable if the starting process of demand planning goes awry. This has a domino effect, adversely impacting all other downstream planning activities.

A combination of factors has made demand planning a significantly complex activity:

• Timely availability of data: Siloed systems and multiple data sources lead to delays in data collection for forecasting, which ultimately impacts the demand plan.
• New product launches: Companies launch new products to meet new customer demands or enter a new segment or for competitive advantage. Accurately forecasting demand for the new products has always been a challenge, especially in the electronics and fashion and CPG industries due to short product lifecycles.
• Chaining customer preferences: Traditional demand forecasting methods rely on historical patterns to generate predictions. When buying preferences keep changing continuously, no single method can generate an accurate forecast.
• Lack of collaboration: Siloed planning and lack of cross-functional communication leads to loss of important information midway, and ultimately results in sub-optimal plans
• Demand interactions: Companies often find it difficult to account for complex interlinkages among product substitutions, promotions, and competitor actions leading to inaccurate demand plans.

An additional layer of complexity is introduced when different methods are applicable for demand forecasts related to different timescales. Typically, forecasts can be classified into three categories:

• Long-term: Generally 2 – 5 years into the future, these forecasts are strategic in nature and are associated with product lines and categories in relation to entering new segments, new markets, etc. These forecasts, in turn, drive mid-term and short-term forecasts.
• Mid-term: With a planning horizon of 6 months to 2 years, mid-term forecasts are tactical in nature and focus on product seasonality.
• Short-term: Typically in the range of 3 to 12 months, short-term forecasts are operational in nature—such as making adjustments and prioritization.

The longer the time horizon of forecasting, the more prone it is to errors. This is because of the uncertainty involved. Hence, even though long-term forecasts drive short-term forecasts, it is important that short-term forecasts are adjusted from time to time to account for the latest trends/insights so that fluctuations can be understood and an accurate forecast is generated.

All of the above-mentioned challenges necessitate employing techniques to analyze a large amount of demand signal data from various sources, in addition to including a number of non-traditional signals which contribute to demand fluctuations (such as social media reputation management).

Adoption of Artificial Intelligence (AI) and Machine Learning (ML) techniques helps solve these complex problems. They deliver a forecast and, in addition, a probability distribution of the anticipated volumes with the help of Predictive models. Probability forecasting is valuable as it provides the means to assess risk and plan for uncertainties.

Predictive models are nothing but a manifestation of the pattern found out from a large amount of historical data which is interlinked, and this helps carry out an exploratory analysis. You could find a number of machine learning methodologies such as Decision Tree Analysis, Feature Engineering, etc., because of the wide variety of applications Machine Learning has in various fields, but we limit our focus on the ones which aid the demand forecasting function. Broadly, these methodologies can be grouped into the following:

1. Regression

Regression analysis is used to explain the relation between a dependent variable and a set of independent variables, by plotting the data values, and finding the best-fit trend line, which can then be used to predict future values. A classic example is the effect of prices on sales; when the price decreases, sales increase and vice-versa, where sales is the dependent variable and price is the independent variable. Further, Regression can be classified into four categories which are applicable in forecasting depending on the demand patterns:

1. Linear Regression
2. Exponential Regression
3. Logarithmic Regression
4. Power Regression

1. Time Series Analysis

Analysis of a series of data points captured in a time-phased manner with equal intervals forms the basis of time series methods. The analyses and further breakdown reveal important components such as Trend, Seasonality, and Periodicity. These components can further be used to extrapolate future occurrences. Following are some of the popular methods which make use of the time series:

1. ARIMA: An acronym for Auto Regressive Integrated Moving Average, these models are capable of describing the time series based on regressing the previous values of itself, leading to an equation that can then be used to calculate the future occurrences. ARIMA models are limited to handling series without seasonality. However, when we come across data patterns that have a seasonal component, as evident in most of the demand planning scenarios, SARIMA can be used, which is Seasonal ARIMA as it supports modeling of the seasonal variations.
2. Exponential Smoothing: Exponential smoothing methods involve generating forecasts with the basic premise that future values will reflect the past, but with the flexibility to assign weights to the past values depending on their recency influence. It involves breaking down the past values to identify the equations: Level, Trend, and Seasonality. These equations are used to extrapolate the data along with assigning the weights (using the corresponding smoothing parameters α, β, and γ). The extrapolation is further added/multiplied (depending on whether it is an additive method or a multiplicative method) to obtain the final forecast.

It is also important to note that, more often than not, there is no single model that can generate a good forecast, it is always a combination of several models depending on what exactly the objective is, available data, and also the set of independent variables underpinning the objective.

Following are some of the approaches/avenues which lead us towards deciding which ML models are to be employed depending on the quality/type of data available.

Demand Sensing

Companies are using advanced forecasting approaches leveraging cutting edge technologies to solve this complex problem. One such approach is Demand Sensing, which considers a variety of factors such as weather, latest market trends, holidays, and other macro-economic indicators, to run predictive analytics, which assists in generating an accurate forecast. Demand sensing greatly improves near-term forecasts by using short-term demand data. Near-term here can be days or even hours, depending on the industry. There are studies that indicate Demand sensing can reduce forecast errors by almost 50% while also improving inventory management.

Competitor Actions

Competitor actions often provide meaningful insights. Nowadays, it is very easy to track the promotions being offered by competitors through web scraping techniques. This data can be used to derive the net effect on demand. Also, situations such as OOS (out-of-stock) at the competing organizations present significant opportunities to augment an organization’s sales.

Forecast Adjustments

Statistical algorithms have their limitations. The benefits of these algorithms are dependent on past sales. Certain scenarios warrant manual interventions by the planners by way of forecast adjustments, as they have the latest information (example: real-time POS order data, weather, business cycles, any catastrophic events, etc.) which can generate a realistic demand plan.

Demand Shaping

In almost every planning cycle organizations are faced with one common question – how to solve the demand-supply imbalance. Either the demand is expected to be more than supply, or vice-versa. When demand is higher than supply, companies resort to increasing the prices in order to shape the demand, since supply cannot be increased in the short term. Other methods to bridge the demand-supply gap could be to augment supply by including substitute items.

Employing a combination of the methods discussed above in conjunction with cutting edge technologies, is the key to presenting meaningful insights, validations, and—more importantly—discovering hidden trends. It also provides demand planners with a clear course of actionable items to overcome uncertainties and execute their demand planning process in a way that is intelligent, all-encompassing, resilient, and which will yield better results.

Recognizing the significance of this supply chain problem, ITC Infotech has built an “Intelligent Planning and Execution” solution powered by our “Platforms of Intelligence” along with our platform partner Anaplan. The solution helps supply chain leaders develop a new view of demand by tracking and analyzing shifts in purchasing behavior, consumer sentiments, and relevant macro causals to better understand the shifts in demand signals over time instead of relying on historical sales alone. Organizations can leverage this enhanced view of the demand for improved supply planning and optimal allocation of resources across the value chain.

A range of supply chain leaders around the world have leveraged bespoke intelligent planning solutions created by ITC Infotech. These solutions have been central to improving their demand sensing capabilities and maximizing commercial opportunities. We helped a leading beverage company improve their revenue by 3.2% in 4 months by streamlining the supply chain and manufacturing using an AI/ML-powered intelligence platform which uses predictive forecasting to generate recommendations – cross-sell, up-sell, sales volumes, orders, etc.  Another leading manufacturing organization saw their forecast accuracy dramatically improve by 25 p.p. in a span of 3 months leading to a 62% reduction in lost sales. To see if our approach and methodologies provide you with a competitive advantage, write to contact.us@itcinfotech.com

Author:
Vinay Varadaraj Mirajkar
Lead Consultant, ITC Infotech India Limited