We are building predictive models for cross channel marketing automation. Our clients use our services to understand the impact of their paid media activities on their business, and to plan the optimal future media activities based on this understanding.
For this purpose, we collect large amounts of relevant data, and build a variety of analytical models. The typical data collected include:
- our client’s marketing activities, particularly paid media,
- competitor’s marketing activities,
- calendar information; holidays, paydays, etc,
- macro-economic factors, such as consumer confidence, unemployment, etc.
Our models predict a predefined KPI - usually some kind of sales parameter such as web store sales, new signups, store revenues, etc.
The models we are building have two major uses:
- To predict the KPI based on a given paid media spend in the past or future.
- To recommend the optimal media allocation - i.e. the distribution of spend per paid media channel, and the temporal distribution over a period of time.
As part of our ongoing research into better marketing models, we want to implement a new approach to modeling which we believe will be more scalable, better performing and easier to maintain. The basic idea is to have several collaborating models, each serving a specific purpose.
With this project, we want to, as a proof of concept, build an "Organic Seasonality Model".
The Organic Seasonality Model is expected to isolate the organic development over time of a client's KPI. One way to achieve this is to remove all other causes of KPI fluctuation, such as the client's own media spend, competitor activity, random events, etc. Other approaches could work as well.
Based on standardized input data, which is the same across all our client, the model should produce a standardized output which can be used by other models or visualized in our client facing UI.
The output of the model is a time series with daily granularity which represents the natural seasonality of the client's KPI. The output should exhibit all seasonal effects, such as:
- The overall trend of the client's KPI; increasing, decreasing or stable
- The effect of day of week
- The effect of the time of year
- The effect of seasonal events (Christmas, Easter, etc).
The model should work for both generating insights, i.e. explaining what the organic past seasonality looked like, and for generating a predicted future seasonality for a given time period.
The latter should take known future seasonal impacts into consideration. For example, if the client's KPI is impacted by Easter, the predicted future seasonality should take the actual future dates of Easter into consideration.
We value engineering skills in the delivered code; a good structure, a reasonable level of software tests, and an attention to the fact that the model should work for a wide range of clients without change.
If the project is a success, we will extend the project to build other models as well.
The Technology Stack
We would like the project to be implemented in Python using our H2O platform which is running on our Hadoop Cluster.
The only restriction on which modeling methodology to use is that it should be supported by H2O. The Experts should choose the methodology believed to be the best.
The project team could consist of 3-4 Experfy Experts who will be working together with a project lead and our Data Engineering and Marketing Science teams.
We are looking for experts with strong capabilities within the area of modeling, data science and software engineering and prefer experts with experience from different modeling methods.
At times during the project we would like the project team to visit and work with people from our Marketing Science teams from Los Angeles and our Engineering teams in Barcelona, Spain and thus we prefer experts who are willing to travel.
Our preference is to have an expert for the occasional On-Site work in Barcelona, Spain, but we are open to considering Experts, who are not able to travel as well.
We are interested in both individuals and already established teams.