How QDT got its start: challenges, learnings and product evolution

QDT got its start building one Machine Learning model for Mars Wrigley to predict cocoa futures. To deploy the first model it took us 6 months of development at a cost of approximately $1million USD. Our technology has come a long way since then, and the results speak for themselves. The following section lays out our solution evolution in terms of:

1. Challenges
2. Features/Benefits
3. Business outcomes
4. What’s next?

‍

1. Challenges‍

A few of the major challenges we faced with our original methodology were:

• Data prep is slow and labour intensive
• Model updates are manual and therefore time-consuming
• There is so much data, and different data is predictive in different contexts
• An individual model is vulnerable to changes in market dynamics

To address these challenges and more, our team developed the beta version of Quantum ML, inspired by some open source code. To make the code effective for time-series predictions we did a significant amount of customization, and, amongst other things, implemented a more sophisticated machine learning methodology, as well as additional analytical tools and visualizations.

‍

2. Features/Benefits

Some examples of developed features and their benefits are:

• Multi-model deployment – allows us to build and run multiple models at once.
• Auto-updates – enables daily updates of models without any human input.
• Group Dashboard (multi-model view) – provides a comprehensive view of multiple model predictions and performance metrics across multiple timeframes.
• Driver Analysis – Quantifies relative impact of specific data inputs across time.
• News Page – Daily news aggregation from 173 news sources with Theme and Keyword filter function to identify impactful news stories.
• Added Algorithms – Expanded library of machine learning algorithms.
• Algorithm improvements – Hyperparameter customization.
• Hyper-optimization protocol – Algorithm optimization protocol.
• Centralized Data Lake – for ease of data access
• Auto-series – Auto-model generation and input optimization.
• Expanded Use-Cases and Timelines

Building off of these new features and the success achieved with Mars Wrigley predicting Cocoa, we expanded the number of target variables to other agricultural commodities.

• Sugar
• Wheat
• Rice
• Corn
• Milk

In contrast to the original cocoa model deployment timeline, building better models for all of these targets took only three months. In addition to doing things faster, we also achieved better results.

‍

3. Business Outcomes

As of today we have worked with the Mars’ Buying Team to build over 4000 models, and have more than tripled Mars’ original cost savings. Current yearly procurement cost reduction ranges from 2-8%, depending on the commodity.

‍

4. What’s next?‍

In late 2020 we expanded the core function of Quantum ML to also provide Revenue and Sales forecasting. We call this Superforecasting. Simply put, what makes Superforecasting so effective is that it incorporates external (global) data inputs like regional GDP, Unemployment, and a long list of additional alternative data types, like Covid case and death rates, that improve accuracy and granularity of analysis.

As of June 2020 we are halfway through our third project with Mars: SuperROI. SuperROI uses the same underlying ML engine as Quantum ML and Superforecasting, but with a different use-case: Marketing Optimization. With SuperROI we have built off of Superforecasting to provide an additional layer of ML and analytics that tells Mars’ how effective their marketing spend is across advertising channels, and how to optimize that spend to maximize sales.

‍