Building Early, Accurate Forecasts of Key Government Reports

Project Description

Context:

We are an early stage tech company focused on combining survey and physical data to create accurate, early forecasts and metrics for a global customer base.

Task:

• A government organization has been in the business of monitoring and forecasting a key economic activity for more than 100 years.  Their monhtly outlook reports have become a critical source of truth to the markets
• The task at hand is to forecast these government reports at both the national and subnational level at least one week ahead of time, refreshed daily.
• Final models will be assessed based on accuracy and earliness across a 10+ year time window (backtesting)

Materials:

• Data are all available at state and national levels on a daily basis for 2003-2016 unless otherwise specified:
• Government  forecasts at the state and national level, at a monthly cadence (this is what is being forecasted in this pilot work)
• Government  end of year actuals at the state and national level (this is what the government is forecasting)
• Survey-based data on a weekly basis that is relevant to the target outcome and available at the state level
• Earth data features at a state and sub-state level for key physical variables; formatted for tabular ingest; hundreds of predictors on a daily basis

Timeframe & Project Plan: 

• Overall Aim for productionalized pilot models by end of July; ideally starting work as soon as possible
• Four week timeline, with assumption of one full time resource
• Weekly milestones and project go-no-go decision at end of week one

Ramp Up | week 1

• 2-3 meetings with the team to understand the domain / challenge
• Data onboarding for all of the materials
• Discussion of workplan
• Data exploration
• Variable selection
• Modeling plan
• Outline of white paper (2-3 pages)
• Very rough, first prototype models (initial results)

Preliminary Models | week 2

• Analysis of early results; backtesting prepared
• Prototype models
• Prioritization for model refinements
• Pre-engineering for putting into production

Model Improvements | week 3

• Revised prototype models
• Updated analysis and backtesting prepared
• Pre-engineering for putting into production
• Draft white paper (2-3 pages)

Model Improvements | week 4

• Final set of changes / permutations
• Placing models into production alongside engineering
• Final QA and backtesting
• White paper finalized (2-3 pages)

Logistics:

Location: Preference for onsite but flexibility for video / remote work; strong preference for roughly in timezone

Engagement: Aim is for full time engagement; 40 hours per week

Tools: R and/or Python in the Scikit-Learn framework; collaborator should be a master of either or both of these frameworks

Modeling: Deep experience in machine learning-based predictive modeling and timeseries; example models where candidate should have long experience with applied work include Random Forest, SVM, Cubist, GBM, etc.

Data engineering: Our team will deliver large, structured data cubes (flat files) for modeling; candidate should be familiar with handling at-scale data challenges; that said, local machine execution should be adequate (no obvious need to distribute or use high performance compute)