Success Cases

Portendo's featured success cases include pricing and merchandising optimization, real-time recommendation algorithms, fraud identification, commercial scoring and trade error detection

Developing a merchandising optimization algorithm for a European e-commerce merchant

Industry

Retail E-commerce

Business Need

Accelerate marketing A/B testing on the various pages of the website acquisition funnel (home page, search page, category page, product page). Marketing used to perform this test by hand, taking 1-4 weeks per A/B test

Complexity

The classical marketing approach is either not robust (marketing extrapolates the differences in performance when it is in fact just noise) or slow (marketing doesn’t optimize the allocated % traffic to each alternative A or B). Accelerating tests requires to manage rigorously, in real-time, the allocated % traffic to each alternative

Data Science Solution

A dynamic exploration algorithm, taking the optimal path between testing each alternative (exploration) and maximizing the P&L with the best alternative (exploitation). It optimizes the allocated % traffic to each alternative in real time depending on the user context and the statistics gathered so far

Sample Insight

The algorithm converges very quickly (snowballing effect on % allocation), freeing time for the next A/B tests. The more alternatives in the A/B tests (e.g., A, B, C, D…), the stronger the snowball effect

Impact

x8 number of marketing A/B tests per quarter

Developing a pricing optimization algorithm for a French e-commerce merchant

Industry

Retail E-commerce

Business Need

Adjust products pricing in real-time beyond simple business rules (e.g., competitors marking and margin objectives) to optimize revenues and P&L

Complexity

The catalog of products is very large (>50.000), its data is poorly structured and lacks semantic information. 50% of revenues are generated by < 5% of them (the “top products”), making statistics hard to gather on the 95% remainder. Seasonal trends (e.g., first spring sunny weekend, Christmas rushes) create business opportunities which are thus hard to detect quickly

Data Science Solution

A dynamic decision algorithm, adjusting, for each item and in real-time, the optimal price path between exploring end-user preferences which are yet unknown (exploration) and maximizing the P&L (exploitation). The algorithm optimizes in parallel 50.000 items, identifying automatically commonalities between items to accelerate learning, and adjusts for seasonal trends

Sample Insight

The algorithm uncovered latent families of product (e.g., specific phone accessories, specific IT equipment) which were not tagged per se in the catalog data structure, and exploited this information to aggregate sufficient statistics to perform optimizations

Impact

+10% conversions, +10% revenues, +5-10% P&L

Developing a real-time product recommendations algorithm for a European advertiser

Industry

Retail E-commerce

Business Need

Identify in real-time user needs & preferences down to the single user level, recommend relevant e-commerce items in a >1 million items inventory, price optimally web advertisements to bring users to the website

Complexity

Traffic data is huge (>100Go per day), product inventory is poorly structured and lack semantic information about sold items. User behaviors are complex, numerous, and detecting each of them requires to capture interactions (“weak signals”) between many predictors simultaneously, e.g., female users jumping from fashion news to clothing websites

Data Science Solution

A suite of machine learning models to identify end-users tastes, cross them with product inventories, estimate their click-through-rates, conversion rates and basket sizes, and manage the ads pricing against competitors

Sample Insight

The algorithm uncovered specific user behaviors (e.g., soccer web surfing, manga web surfing, newsite reading before jumping to an e-commerce site) which drove excellent click statistics but near-zero end-of-chain conversions. It cut them, saving expenses

Impact

Fully automatic channel participating to >200.000 real-time ad auctions per second, top 10 Google ads purchaser in France, managing >50 million euros yearly online sales

Developing a fraud identification algorithm for a European consumer lender

Industry

Financial Institutions

Business Need

Detect fake credit applications before application review teams, to optimize the operational and credit cost of fraud. Fraud and fraud detection operations are, with default and sales operations, one of the 3 drivers of P&L costs for a consumer lender

Complexity

Frauders try to disguise their credit applications as perfect “sons-in-law”, circumventing simple checks. They may also use stolen ids, real employer HR phone number (through an accomplice) or fake addresses, rendering these deeper checks ineffective

Data Science Solution

A machine learning model to learn the application patterns of frauders and detect them in real-time

Sample Insight

Divorced single women with children and an existing consumer credit had excellent fraud scores (<2%), counter-balancing their low/medium credit score

Impact

Model detects from 15 to 30% of frauds before any human analysis, freeing FTEs to accommodate business growth

Developing a commercial scoring algorithm for a European consumer lender

Industry

Financial Institutions

Business Need

Maximize sales performance by prioritizing customers to be called by the outgoing calls sales desk. The sales & application process is a major source of P&L costs – targeting the highest probability customers allows to reduce both the sales costs (fewer callers needed to reach a target volume of applications) and the application process costs (lower abandon rate on partial applications)

Complexity

Sales desk managers (operational background) rely on their “soft” experience to support unclear, under-performing, prioritization rules. The platform growth changes the mix of consumer profiles, outside managers experience. Lastly, the predictors of a consumer willingness to complete an application are multiple (>30 markers of interest) and interact with each other (“weak signals”): e.g., an applicant with good revenues financing some home improvements won’t behave the same if he is single or married

Data Science Solution

A statistical & machine learning analysis identified a subset of client criteria - which didn’t fit the operational managers intuition - driving >90% of the success of a sales call

Sample Insight

Weak signals existed and brought a useful performance increase, but the most important action was, against the operational managers experience, to drop all the existing rules and sort by the most important criteria: deal size

Impact

The most simple variant of the model generated +25% sales per outgoing call

Developing a trade error detection algorithm for a leading global bank

Industry

Financial Institutions

Business Need

Catch immediately market trade ticket errors before they are processed by middle & back-offices, as they generate operational work and counterparty risk & fees

Complexity

Simple checks have already been practiced for years by the teams, and cannot be extended as the bank activities are incredibly diverse : e.g., most major currencies are used, with widely diverse ticket amounts. Additionally, the trading business undergoes frequent changes, which are not anomalies

Data Science Solution

A machine learning model trained to identify errors based on past history, coupled with a statistical model to detect new trade configurations (novelty) which have no past antecedents

Sample Insight

Anomaly patterns found are complex, often requiring the simultaneous interactions of several predictors (e.g., dates ordering, amounts, currency, trader involved)

Impact

>80% errors catched immediately, with <20% false alarm rate. Middle office STP metrics improved five-fold under same costs

Lessons learnt to ensure success of Machine Learning initiatives

  1. It has to be important. The value at stake must be significant enough for the firm to allocate priorities and resources to ensure success
    • Value must be visible to senior management to ensure priorities are set accordingly
    • Necessary resources: IT, Operations and business managers must be committed to the project
  2. Go for a business that is reasonably stable. Ensure you do not have too many “moving parts”
    • Strategy and priorities are clear
    • Operations are stable
    • IT platforms function at an acceptable level
  3. Ensure you have the right mixture of the right talents
    • Management and business know-how and experience
    • Data science experts with experience in the business area
    • Use a performance focused methodology (efficient and effective)