The science behind Sprint AI

Powered by Data. Empowered by intelligence.

Dynamic demand forecasting using Machine Learning, coupled with advanced optimization techniques to revolutionize inventory and assortment decisions.

Self-learning algorithms that can handle scale of thousands of SKUs and large store & warehouse networks.

Powered by Data. Empowered by intelligence.

Everyday decisions to be made by retailers & brands

Lifestyle retailers and brands make thousands of inventory movement decisions on a daily basis.
  • Which products do I stock at a store?
  • How much stock should I allocate to stores as my first drop?
  • When & where should I move products from warehouse?
  • Should I move products out from a store if they are not selling & if so where to send them?
  • Should I fulfil online orders from a store or a warehouse?
  • Which store should I fulfil the online order from?
  • Which styles are trending?
  • Should I re-order, how much?
These are fairly complex and involved decisions and take a lot of time and effort to solve optimally...

Using past data has always been challenging in lifestyle retail…

  • Using data has always been challenging in lifestyle domain. Data is often sparse and there is a lot of noise and volatility.
  • Further, the historical data has a lot of rich information, but it has to be normalized for multiple factors (promotions, events, weather, option availability, channel growth, new category introduction, etc.) for it to be relevant in the future. Often a lot of these normalization are non-trivial and require one to segregate the impact of different factor

Failure of rule-based decision making systems

Almost all traditional demand forecasting techniques perform sub-optimally in the lifestyle domain. There needs to be context aware (selection, availability, promotions, seasonality, events, etc.) approach to forecasting which normalizes for the past context and forecasts for a future context. Further, you need algorithms that can work well for sparse data sets as quite often you have limited data and you need an ability to filter out noise from signal.

Additionally, inventory movement decisions are connected - store allocations influence replenishment, inter store (or inter warehouse) movements and online fulfilment from stores. There are trade-offs to be done across these decisions (over product life cycle) and they need to be optimized together. Often you don’t know where the improvement opportunity lies and it’s very hard to solve these problems simultaneously. Retailers end up applying heuristics and rule-based approaches to solve these problems which are either sub-optimal or time consuming.

The limitations of out-dated forecasting techniques and rule-based decision making for optimization will not suffice for the future we are entering.

Enter Sprint AI Science

Artificial Intelligence and Machine Learning, combined with advanced optimization techniques like reinforcement learning and dynamic programming are poised to change the way we will make inventory movement decisions.

At SprintAI, we leverage ensemble of machine learning algorithms to sift through multiple data sets to better understand your future demand streams across your network. Unlike traditional techniques, these algorithms don’t depend on linear relationship between variables and can un-cover non-linear impact of different factors in tandem.

We then run millions of possible scenarios over the entire product life cycle with different stocking levels and inventory & fulfillment policies to arrive at the most optimal product decisions to be done at a point of time.  

Our Holistic Approach to Inventory and Fulfilment Optimization

We have a two-pronged approach to making inventory movement and fulfilment decisions:
1. Granular (Store X Product level) and accurate forecasting of future demand across both online and offline channels based on future context
2. Real-time optimized inventory movement or fulfilment decisions at the current moment, considering the entire product life cycle across the network

Granular (Store X Product level) and accurate forecasting of future demand across both online and offline channels based on future context

  • Ensemble of machine learning models like Decision Trees, LSTM, Time Series Forecasts and Spectral Forecasts that learn from inter dependencies between different types of features. These algorithms work at different levels of hierarchy and improve the accuracy of predictions. We have built a proprietary challenger selection system for selecting different algorithms for different class of products.  
  • Context aware models that normalize for the past sales of the product based on the context (product availability, extent of promotion, assortment available, seasonality and new introductions) and then use that data for modelling the future in which the context may be different.  
  • Dynamic and responsive probabilistic models that keep updating based on new data on daily basis. We understand that the demand is not deterministic. Rather than looking at expected demand, our systems look at our understanding of the future probability distribution of demand (Eg. We model the probability of selling 2 or 3 units of a SKU in the next 1 week, rather than the expected sale of a SKU next 1 week).
  • Our models work well in case of sparse data sets as well, as the learning happens not just based on the product’s rate of sale at the store (often misleading, especially in the first few weeks of launch). Rather than that, we also learn from the performance of product across stores, performance of similar products at the store and many such factors.  
  • Our models also understand where (cluster of pin-codes) the demand is coming for online orders. Before modelling the future demand for a store, it’s important for our systems to understand where has the past demand come from. Has it come from the same store’s catchment, from another store’s catchment or coming from a completely new customer?

Real-time optimized inventory movement or fulfilment decisions at the current moment, considering the entire product life cycle across the network

  • Local decisions of inventory movement or fulfilment are driven by simulating out the complete product life cycle and arriving at the most optimal decision at the current point of time. (Eg. Only send higher first allocation quantity to stores if it is leading to lower chances of overall out of stock levels across the season).
  • Inter-temporal optimization techniques (which are widely used in algorithmic trading) are used for optimization that work for cases where the future demand is probabilistic and uncertain.
  • Specialized Optimization Solutions that leverage dynamic programming and reinforcement learning. Example of few such optimization use-cases include
    1. Most optimal product movements for inter store or inter warehouse movements to give the maximum revenue up-side with minimal logistics cost
    2. Replenishment quantity optimization in cases of limited quantity left in the warehouses or limited space at the store
    3. Trade-off between future markdown versus fulfilling an online order from a store that might have a slightly higher transportation cost
    4. Trade-off between width reduction at a store and the revenue up-side from sending the same product to another store.