Digital Marketing: Campaign Performance Boost Based on Weather Data

Digital marketing campaign impact strongly depends on user behavior, which is directly affected by weather conditions and social events. This is particularly sensitive for campaigns targeted to position ads in the results of keyword-based users searches within platforms like Google and Bing.   In this scenario, JOT Internet Media has developed, in collaboration with EW-Shopp technical partners, “Campaign Booster” a new service integrating weather forecast, events, and marketing performance data to predict the performance of campaigns for categories of keywords. This enables account managers to predict the adequate date to launch new campaigns depending on the keyword-weather variable/event correlation, maximizing the number of impressions (as key impact factor).

The need for a more automated approach to enrich the data with the aim of optimizing marketing campaigns is crucial for a company like JOT, which runs campaigns in 79 countries, 17 languages, collecting 15TB data about performance KPIs. 

WATCH THE VIDEO: Detecting user behavior patterns to boost digital campaigns

KEY ASSETS

Key assets of the new service are focused on the enrichment of the data value chain by:

  1. Integrating external data (weather and events) with marketing performance indicators
  2. Aggregating the data sets according to the expected campaign management strategy
  3. Generating temporal models about the keywords/categories behavior
  4. Implementing 4 different predicting services supporting a more efficient digital marketing campaign management

THE APPLICATION

The core of Campaign Booster is in its data enrichment and analytics components. These components enrich JOT campaign performance data with weather data to build predictive models. The models are used at run-time to estimate the performance of the campaigns using fresh weather forecasts for the incoming weak.

In an early version of Campaign Booster, described in a paper presented at ISWC2019 (an open version of the paper can be found here), the service tried to estimate impressions for individual keywords, building a model for each of it. However, in order to scale-up predictive modeling – JOT has data about ~4B keywords – and to overcome the sparsity and weakness of signals provided in these data, a newer approach has been set up, where models are built at the category level. Categories are associated with keyword-specific campaigns. The idea was therefore to build representative signals for categories in the third level of Google’s category taxonomy, to use more reliable signals to learn predictive models.     

Based on this idea processes have been set up to: 

  • Build representative keyword clusters for categories, by finding the keywords that are more semantically related to the category labels;
  • Enrich (historical) performance data at keyword-level with their associated category – if part of a representative cluster, and (historical) weather data; this is done by using the EW-Shopp data enrichment tools
  • Aggregate keywords’ impressions at category and region level, e.g., impressions for the Car category in the Bavaria region, as impressions are sensitive to both categories and regions;
  • Train a machine learning model to predict the impressions for a given category in a given region, e.g., for Car in Bavaria. 

When an account manager has to launch a campaign for an input set of keywords, these keywords are associated with their most similar category. Weather forecasts for the incoming week are collected and the model for the associated category is applied to obtain the predicted impressions for each region and recommend the best-estimated date for launching the campaign in each region.

Estimating category-level and keyword-level impressions is nearly impossible due to the large amount of indicators involved, including user behavior random variability. Impressions depend on the specific keywords being activated, the investment, and several external factors that are not tracked in the data, like ad platform algorithm updates. However, category-based estimations (aggregating keyword-level data) and the consideration of important factors such as weather and seasonality (implicitly learned from historical data) at a large scale, i.e., for several campaigns based on large numbers of keywords, provide a significant impact boost to the optimization of campaign management with visible results on the overall monetization (see achieved impact).    

ACHIEVED IMPACT

JOT has improved the most relevant marketing indicators in digital marketing such as impressions, clicks, and CTR. All these are related to both the quality of the campaign management and the traffic generated for the client’s site. High-quality traffic generates some kind of conversion on the landing page like sale, lead, or expression of interest. The achieved business outcomes can be summarized as follows: 

  • Time to get impressions has been reduced from 1 month to 1 day
  • Initial impressions rate has increased up to 30-50% (average value 5-10%)
  • Lists with a smaller number of keywords are now used, thus leading to concentrate the investment on relevant keywords only

INDUSTRY IMPACTED

Campaign Booster mostly impacts on JOT’s campaigns, generating high qualified traffic to our clients’ sites.

In the future, this service can be used to boost the impact and performance of digital marketing campaigns of SMEs, where the keyword list describing their business and service are shorter and the influence of the nearby conditions is more critical.

RESEARCH & INNOVATION DOMAIN: Data integration, data aggregation, modelling, prediction, user interface and simple visualization

TOOLS INVOLVED: Google API, ArangoDB, Grafetizer, Python based models and java script node runner

CONTACT PERSON: Fernando Perales fernando.perales@jot-im.com