"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

May 11, 2021

Price Optimisation: From Exploration to Productionising

Key Notes
  • Optimize for revenue / profit / market share
  • Lose market share but retain profits
  • Lose price but increase market share

  • Standardize - Learn from business, Apply Algorithm
  • Apply / Learn, Avoid fluctuations, inconsistencies
  • Optimize - Know both win and lost scenarios
  • For loss, how price can be pushed
  • Explore - No more data to learn from
  • AI solution to explore / Reinforcement learning problem
  • Learn by itself (Rare occurrence)
  • Working model - API
  • Hosted model
  • Web front end
  • Building products
  • Learned from historical data
  • Business knowledge
  • Existing Systems
  • Sales Data
  • Market Research
  • Features - Pricing
  • Perishability
  • Premiumness
  • Cost
  • Frequency of Sale
  • Seasonality
  • Market features

  • Feature Implementation
  • Clustering, Encoding, Algorithms
  • Algo choices
  • Segmentation
  • Segment dataset
  • Data Driven Unsupervised method


Break things earlier if it saves us

Segmentation - Unsupervised vs Business Driven
  • Business driven explainable
  • Ownership

Category driven models - encoding

  • One hot - Yes / No
  • Factors - Assign numerical value, Ordering issues
  • Binary - Introduce hierarchy, Reduce objects to describe objects

Base Algorithms


Depending on the time-series aspect

Code Optimization

POC mentality

Changes in the market - Time series and frequent retraining


Make model more time-dependent, time-series



Short term corrections



  • LSTM to output median difference
  • Different between regression model and error calculation
  • Rolling average based correction
Production implementation
  • Azure, Flask, Kubernetes

  • Kubernetes deployed models
  • Load balancers - Containers

Data feeds - Landing zone - Data Engineering - Consistency Check - Databricks layer 
  • API End point - Model in Flask - Data Pipeline
  • Model confidence



Optimization methods
  • Optimizing without data
  • Price elasticity


  • Use bands
  • Historical data
  • Quantiles
  • Lower and upper bands





Elasticity
Price elasticity vs Sales forecast
Historical data for every value with respect to price
Pseudo cells, pdf distribution






Keep Thinking!!!

No comments: