1st PyData Piraeus: Blueground 際際滷s - How Machine Learning is Reshaping Pricing
0 likes407 views
Pricing is a complex optimization problem requiring multiple data inputs. In particular, the real estate industry has vast data that is often fragmented. Companies in the space, such as hotels and investors, have realized the power that real estate data holds. At Blueground, we are at the early stages of developing a predictive model to optimize our pricing. In particular, we are using internal and external signals to predict whether our units will get booked and ultimately give price suggestions. We recently started developing our pricing model and are excited to share what we have learned from this journey so far
1st PyData Piraeus: Blueground 際際滷s - How Machine Learning is Reshaping Pricing
- 1. Show up. Start living. PyData Piraeus, How Machine Learning is Reshaping Pricing September 27th, 2019
- 2. D A T A & B I BLUEGROUND BI TEAM AND AREAS OF FOCUS Strategy Data & BI Growth People Ops Primary focus areas Pricing Asset selection Sales Channel & Mkg BD CX Competition Analytics and visualizations City and individual KPIs
- 3. Define Goals D A T A & B I HOW ML IS RESHAPING PRICING OPTIMIZATION Pricing Challenges Changing market conditions Different products Understanding consumer behavior Collect Data Train Model Price Recommendation Quantity & Fluctuation of Data Pricing Optimization Process
- 4. Q3-Q4 19 onwards Model-Driven & Automated D A T A & B I BLUEGROUND PRICING APPROACH Q1-Q2 19 Analytics & Rules-Driven Q3-Q4 18 Human-driven & Manual Initial price Probability to book Price recoms. Price sync PRICING MODEL Optimization Goal ≒ Maximize Price ≒ Minimize Vacancy
- 5. Number of inquiries BG unit availability Number of bookings Website traffic Deal duration & source D A T A & B I DYNAMIC PRICING DATA INPUTS Market Data City Data Unit Data Lead Data
- 6. D A T A & B I FIRST MODEL ITERATION (LEAD LEVEL) Model Specifications Models: XGBoost, light GBM, GBM Advantage: Focused on computational speed, model performance and interpretability Features: ~80 features organised in 4 main clusters Interpretability: Elasticity of each lead to selected features Goal: Predict probability of sales inquiry converting to booking
- 7. D A T A & B I INTERPRETABILITY
- 8. D A T A & B I OUTPUT
- 9. D A T A & B I LEARNINGS Garbage In = Garbage Out Collect Data Early Productionize Early Invest in a Solid Testing & Production Environment
- 10. D A T A & B I SECOND MODEL ITERATION (AGGREGATE LEVEL) Model Specifications Models: Combination of classification and regression Features: Added features around market, availability, channel presence, top funnel Goal: Predict probability of unit getting booked by day of vacancy & expected date
- 11. D A T A & B I OUTPUT & NEXT STEPS Next Steps New features Prediction interval Price elasticity A/B testing Assess model predictions vs. user decisions
- 12. D A T A & B I QUESTIONS & FOLLOW-UP Thank you! Questions? Were always looking for new ideas and to grow our team. Contact: Yannis Sarantis - yanniss@theblueground.com Konstantinos Karagiannis - konstantinosk@theblueground.com Leonidas Tsolas - leonidas.tsolas@theblueground.com