Tools & Technology

Algorithms in Adaptive Planning

A time of big change for FMCG

The FMCG industry is seeing major changes as large FMCG companies are being taken on by highly innovative, nimble, ultra-focussed microbrands. The days of blind brand loyalty are over. Today, rather than buying into brands based on reputation, customer demand is based on how quickly and conveniently they can get their hands on the products they want, how much they identify with the products and how deeply their products understand them and personalize to their needs and desires.

In our blog on adaptive planning, we discussed the march of events leading up to the need for the sense and pivot paradigm for planning, and how Adaptive Planning delivers on this vision.  To quickly summarize - the FMCG industry is in the midst of a full disruption, with large FMCG companies are losing market share. Competition is fierce and as a result, operational complexity has spiked – inundating the traditional planning machinery. Planning is turning into an increasingly uphill task for planning organizations; planners cannot humanly keep up with the granularity and effort needed.

Companies recognize these problems but are wary of the hype surrounding new planning technologies. In this article, we answer two questions:
1. Why incremental approaches, legacy systems, etc. will not work going forward
2. How AI and Cloud Computing work to deliver Adaptive Planning

What doesn’t work, and why

CPG companies who want to survive the disruption must think about their ultimate goal — and then build a strategy to support that goal, rather than doing things that are convenient and incremental extensions. Before we look at why we need AI and Cloud Computing to deliver Adaptive Planning, let us take a quick look at what doesn’t work.

Legacy systems don’t work for a reason – they are no longer contextually valid.  Legacy systems that you see today were designed for business conditions several decades ago; they simply lack the speed and scalability required for modern FMCG planning.  Most so-called Advanced Planning and Scheduling systems show effects from this age – they are essentially spreadsheets in disguise. By design, legacy solutions expect lots of manual hand-holding and attention, and often require specialized dedicated admin and IT resources. Finally, they do not pay enough attention to the cognitive burden they offload on planners. For example, they often run several forecasting algorithms and offer the planner a choice of which algorithm to use in practice, instead of making this choice for the planner. They also often expect planners and admins to fine tune forecasts on an ongoing basis.

Classical Algorithms work on averages, and with a limited number of variables and constraints - a carryover from a time when average based planning was state of art. Moreover, they make several statistically simplifying assumptions to model the real world, instead of being truly data driven. They expect data to be near-perfect to produce good results, something that even the best companies struggle with. As a result, these algorithms perform poorly under current conditions of high volatility and uncertainty.

Do-it-yourself (DIY) Sometimes, companies swing to the other end of the spectrum and start building their own planning systems by combining data science and technologies. Very few companies have the technical teams to pull this off – the result is a proliferation of tools and technologies woven together into complex, irreproducible workflows; in short, a disaster.

Modern Adaptive Planning

Modern adaptive planning recognizes that business is not business as usual.  Planners today need to deal with 10X or even 100X the number of combinations their counterparts dealt with a decade ago - it understands and thrives under uncertainty. Adaptive Planning prioritizes speed and automation, e.g. even though billions of calculations are involved, our cloud solution is engineered to be massively parallel - it dynamically scales to large workloads and completes creating numerous planning scenarios in just a few minutes. Here are five capabilities that illustrate how Adaptive Planning differs from today’s traditional planning, and what technology and algorithmic advances are powering it at Algoshelf.

1. Accurate, probabilistic, short-term demand forecasts

A machine learning algorithm creates granular daily forecasts based numerous demand signals including historical sales, promo campaigns, store open/close days, weather, holidays, etc. AI algorithms use 5-10X more demand signals as inputs.

Competing solutions deliver median forecasts, which have  two problems - (1) These forecast target a 50% probability of future demand; however, brands seek high shelf availability and service levels, typically in the 80-95% range, and (2) With a single median forecast, it becomes impossible to make revenue-inventory trade-offs over a range of possibilities, scientifically.

Our adaptive forecasting engine estimates the probability distribution of sku-location-day forecasts. Once this distribution of forecast is known, other algorithms can plan for the optimization of subsequent inventory-replenishment decisions. All this comes at a large computational burden, which is handled efficiently and economically by massively parallel processes on the cloud.

2. Data Driven New Product Forecasting

New Product Introduction (NPI) is a significant, increasingly frequent activity for most FMCG companies. Competing solutions cannot forecast new products, because of lack of sales history. At best, they can create forecast from fake history (pseudo-history) and for this, they make planners go through enormous manual effort of identifying and chaining histories of like store-sku combinations, etc. The burden of all the analytics required to get it right is shifted to the planner, and the planning system essentially becomes a glorified spreadsheet.

In contrast, out Adaptive Planning solution recognizes the wealth of data in your existing sales history and mines this data to extract relationships for new product forecasting – it uses machine learning algorithms to automatically learn patterns from various like-intersections and like-attributes e.g. like brand, like category, same city, same store-category, etc. from different slices of your sales data, and uses these to forecast new product sales for the first few weeks (i.e. until the products establish their own sales histories)

Doing this successfully requires automated feature engineering at scale, and continuous, automatic training of machine learning models for NPI forecasting, and all the burdensome computations are handled by the cloud.  All this complexity is hidden away from planners, so that they can focus on their core task – planning!

3.  Operational Planning in Meticulous Detail

FMCG operations planning is a complex task. It involves tens of thousands of store-location combinations, along with the specifics of daily, weekly, monthly seasonality patterns, the specific promotions being run, the attributes of the store/location such as the specific store replenishment days in the week, maximum shelf space available, the attributes of the products including the different lots and their individual expiry dates, etc. at each sku-location

Competing solutions merely create aggregate forecasts, and simply ignore many of these supply chain constraints, assuming that they will not make a big difference overall. For example, they may treat lead times as constant, ignore shelf life considerations, assume that demand is static, etc. They leave the difficult task of creating the inventory and replenishment plan to support the demand plan, and checking all the meticulous reality related details, to the planner.

In contrast, we feed probabilistic forecasts into dynamic simulation models and fully extract the complete spectrum of probabilistic revenue-inventory relations (trade-off curves) subject to every single one of these operational constraints. This in turn paves the way for the scenario-based optimization of inventory-replenishment decisions. Running such massive simulation routines is well beyond the capability of traditional systems; we accomplish this by distributing the simulations, model training etc. dynamically using on demand cloud CPU and GPU resources.

4. Scenario Based Revenue-Inventory Optimization

Planning attempts to simultaneously satisfy multiple objectives and align activity across sales, marketing, finance, and manufacturing / distribution. Doing this well across tens of thousands of sku-location combinations is non-trivial.

When the task of working out the details is forced on to planners, they tend to live in the weeds, spend most of their time juggling spreadsheets Detail dysfunction sets in - a myopic focus on the business details results in planners and managers losing sight of the big picture. When planners use adaptive planning, they step away from creating detailed forecasts, calculating inventory quantities, etc. and instead play the role of an orchestrator. They assign / revise KPI targets and constraints for each Segment. Once plans are run, they review exceptions and improvement opportunities and discuss these with their peers in sales, marketing, finance, and manufacturing.

Our optimization technology works on maximizing segment revenue subject to a rich set of constraints on service levels, return percentages, working capital available for deployment, and minimum acceptable service levels at a granular sku-location level. The optimization can also be extended across Segments, to a Planning Entity level.

5. Knowledge and Cognitive Analytics

Back to the topic of how legacy systems force the task of working out the details on to planners, making them spend most of their time juggling spreadsheets – nowhere is this more evident than in munging data and extracting useful information and insights

The productivity loss to the organization from repeated, redundant, error prone analysis cannot be overstated. Moreover, most analysis done this way is dependent on the skill of the person, slow, and in most cases irreproducible and untraceable. Even when reporting or BI technology is involved, the task boils down to the user storing data in tables and running procedures on them.  Unless planners can find the insights that they need quickly, they cannot be effective.

Our AI approach replaces all manual work by extracting knowledge out of data and storing it in a Knowledge Graph that understands your business. Machine learning algorithms are used to interrogate this graph to reveal patterns directly to users. E.g. to answer a question such as ‘Which Brands show Winter Seasonality in the East Region’, or ‘How does Segment X compare to Segment Y over the last 6-Months with respect to returns %’, we do not expect the planner to do any special workflow. Just ask the question and get the answer – Cognitive AI will work with the Knowledge Graph figure out how to find the right data set, build the right intersections, conduct the right analysis, and present the answer to the planner in human understandable terms.

Focus on FMCG Planner and Planning

Having worked for years on implementing enterprise systems, we saw first-hand how traditional enterprise systems cared more about offering a plethora of features than on how planners would use them to realize benefits (see our article on enterprise UX). We also realized that the best technologies are only as effective as the buy-in users have in them. We explicitly designed our adaptive planning solution ground up to minimize the cognitive load on the planner, and maximize speed and clarity in decision making. In Algoshelf, all the powerful AI algorithms, Cloud Computing etc. are wrapped in a beautiful UI that hides all non-essential detail and unnecessary functionality from the planner.

Make it Happen

FMCG companies who want to survive this disruption revolution must think about their ultimate goal, and then build a strategy to support it. Companies are looking to their executive team to create a sustainable advantage to survive the changing tides. What is needed today is bold action.

Thanks to advances in AI and cloud computing, adaptive planning is now practical, achievable, and affordable. These tools and processes need to be in place to contextualize and transform planning to deliver this sustainable advantage. The imperative to be more agile is too pressing to be overly cautious – don’t wait for perfect data, systems, etc. By embracing adaptive planning, companies can turn volatility and unpredictability into sources of competitive advantage. In the process, you will unshackle your data, unshackle your planners, and unleash your organizations’ full operations potential