ADA

Achieve forecasting accuracy with IA’s AI-guided forecasting Model ADA.

Our state of the art COVID forecasting models are
achieving high accuracy during COVID-19

Rapid assembly of emerging trends from recovery

Traditional models are slow to factor in new and emerging patterns. We feed critical external data
in our models to train them on region and sector wise COVID severity, as well as
choose algorithms best suited for fast patten recognitions

Recency

Our models are assembled with the capability of rapidly re-evaluating coefficients of key business drivers and are quick to adapt in dynamic environments

Innovative external data sources

Innovative and current data sources such as Google and Apple Mobility Reports, region wise Unemployment data and COVID severity indices help us set granular region level forecasts

Hierarchical forecasting

Hierarchical forecasting allows us to generate SKU Store level forecasts with high accuracy

Updated demand models automatically recalibrated
COVID-19 risks and recovery

COVID-19 risk is imputed based on
multiple external factors

Google Mobility Retail Drop % at County Level

Unemployment Claims data from the Federal Reserve

State-level regulations and COVID-19 incidences

Demand models that weigh in recency and locational COVID-19 impact will be the key to accurate forecasting

Rubber Duck Forecast Adjustment

– ADA works equally well in volatile circumstances.

Such adjustments incorporated into the algorithm equip them to take care of future disruptions – both positive and negative
Such adjustments incorporated into the algorithm equip them to take care of future disruptions – both positive and negative

What do AI forecasting algorithms bring to the table?

Fast pattern recognition

  • Algorithms break down outcomes by drivers.
  • Forecasting algorithms break down final quantity into seasonality, trend, promotional events, etc
  • Any changes in drivers can be detected by algorithm and impact on outcome thus predicted

Process internal and external data

  • AI Algorithms can process data from internal and external sources and calculate impact on outcome.
  • For instance, how does unemployment rate impact business for an organic grocery store vs a dollar store?

What-if scenario analyses

  • Once you have the final quantity as a function of driver variables, you can do any number scenario analyses

The X-factors that makes ADA’s forecasting accurate and superior

Deep Retail Business Context

  • Accommodates for multiple factors and inputs that makes the models very specific and nuanced to retail business
    • Seasonality
    • Lost Sales
    • Marketing / Assortment Spends
    • Trends / Fashion Judgment
    • etc…
  • Embodies learning across a vast set of real-world retail data sets

Superior Mathematical Engine

  • Using an ensemble of AI algorithms to design models, each model optimizes from among 15K+ possible constructs
  • +30 parameters on which models are built compared to standard models at 3
  • Trained on +2 PB of data
  • Trained for +5 years and continuously being refreshed
  • Tested and adjusted for events like COVID-19 through a robust rubber duck / regime change
    construct

IA’s framework is deeply adapted to the business context of retail

Robust retail contextual treatments of the
data before and after modeling

Estimates lost sales and performs business specific outlier removal

Validates model elasticity with business judgement and look back simulations
Creates context-based variables to handle seasonal assortment, inventory changes and new launches
Incorporates trends, seasonality, events, weather, marketing spends, panel data and store closures
Variable transformations (Log, box-cox etc.) based on sales patterns of the business

Framework embodies the learnings across a
large number of retailers

Identifies sales lifts due pricing, promo and coupon changes separately
Caters to regime changes in businesses (e.g. Assortment changes, COVID impact, Store closures, New SKUs etc.)
Tailored to provide recommendations at any level of geography, product hierarchy & time

Recommends promo types (% OFF, BxGy, Price point etc.) suitable to the business

Handles business specific product/customer groups (e.g Fashion vs Core products, Discount seekers vs Regular customers etc.)

Modelling problem framed according to business context

Key considerations Common Retail Application
Buy/Assortment Decision Allocation Promotion (Direct Mail)
At what level should you forecast?
Optimal level Choice countXStore SKUXStoreXWeek SKUXChainXWeek (Assuming chain level promotions)
Why is this optimal? While financial plans (MFPs) are made at a department-chain level, actual buy orders are at a Choice count level. If the customer’s first choice style or brand is not available, they may buy something else. But if they are not happy with that purchase, there is a risk of losing the customer or that part of their basket forever Even if promotions are planned at a “promoted product group” level, it is important that the forecast be available at an item level for supply chain to be able to respond correctly. Promotions are typically the same across the chain, so a chain level forecast is appropriate for the business problem. It will also be more accurate than a store level forecast
How far out should you forecast?
Typical lead time 6-9 months/depending on supply chain agility Weekly/depending on supply chain agility 3/4 months out/depending on marketing collateral deadlines
Why is this typical? Buying for the next year is typically done 9-12 months before the start of the season. If your buying process is different and requires more/less lead time, ensure that the model choices reflect that Replenishment from the warehouse to stores is done every week. If your replenishment frequency is higher/lower, ensure that the model choices reflect that. Typically marketing print formats are frozen 3-4 months before the beginning of big promotional events. If your promotional process is different, ensure that the model choices reflect that.

What should the model be able to explain?

Key Drivers Attributes, Size profile Seasonality, lost sales Promo Price Elasticity, Seasonality
Why are these drivers important? Attributes, Size profile An understanding of seasonality helps you smoothen out the effect of peaks and troughs in demand. Understanding lost sales is helpful in correcting mistakes in the current decision-making process and errors of one time period are not perpetuated to the next period Understanding of promo price elasticity and seasonality helps you stay accurate in forecast even when the offer changes from LY. Otherwise you are either left with inaccurate forecasts or same offers as last year.

Common challenges

While plans are made at a department level – actual buy orders are at a SKU level Incorporating lost sales without over/understating it Same offers are repeated every year; Estimating elasticity of new products

IA has developed robust techniques for improving model accuracy and business interpretability

Raw model outputs

Model type Error from model Elasticity coefficient from model
Linear – Linear 10% 1562
(Interpretation Not sure)
Log – Linear 15% 1.2
(Interpretation Not sure)
Log – Log 12% 4
(Elastic)

Different models provide various elasticity coefficients making it difficult to benchmark models against each other

.Best model based on lowest error

IA’s tuned model comparison

Model type Error from model Elasticity coefficient from model
Linear – Linear 12% 0.1
(Inelastic)
Log – Linear 14% 3.7
(Elastic)
Log – Log 12% 4
(Elastic)

IA benchmarks outputs from each algorithm to ensure that a fair comparison in made before arriving at the right estimates for elasticity

.Overall best model based on error and interpretation

IA’s AI models are built by choosing from a vast
array of possible constructs helping embody deep complexity

Possible different constructs explored by each model is 15K+ and best combination chosen

IA ensures automated forecast model works equally well in volatile circumstances – Rubber Duck Forecast Adjustment

Such adjustments incorporated into the algorithm equip them to take care of future disruptions – both positive and negative

Prevents radical deviation of forecasts and actual that makes it easier to adjust to disruptions

Though, a continuous human A.I collaboration is still necessary to take care of any other unforeseen situations

Our robust forecasting model ADA powers all our SaaS products

Overview

“It is very hard to predict, especially the future’ – Neils Bohr

Every industry uses forecasting while making decisions.

In 2021, traditional forecasting methods, that lean heavily on history, will not work.

Business is changing fast and you need an AI forecasting solution to keep up with the pace.

IA’s AI-guided forecasting model “ADA” has been developed over the last 5 years, handled more than 2 PB of training data and are able to evaluate your data through 15000+ constructs before generating a business driver-based model for your problem

Success Stories

ADA

Let ADA helps you to get your forecasts right.

Schedule a demo with us!