Improving official statistics in emerging markets using machine learning and mobile phone data

EPJ Data Science

Table 1 All 31 indicator families computed by Bandicoot divided by the temporal unit of the distribution they measure

Temporal period	Indicator families
Over weeks	active_days__callandtext
	number_of_contacts__text
	number_of_contacts__call
	percent_nocturnal__text
	percent_nocturnal__call
	percent_initiated_conversations__callandtext
	percent_initiated_interactions__call
	response_rate_text__callandtext
	entropy_of_contacts__text
	entropy_of_contacts__call
	percent_pareto_interactions__text
	percent_pareto_interactions__call
	number_of_interactions__text
	number_of_interactions__call
	number_of_interaction_in__text
	number_of_interaction_in__call
	number_of_interaction_out__text
	number_of_interaction_out__call
	number_of_antenna
	entropy_of_antennas
	percent_at_home
	radius_of_gyration
	frequent_antennas
Within weeks	call_duration__call
	response_delay_text__callandtext
	balance_of_contacts__text
	balance_of_contacts__call
	interactions_per_contact__text
	interactions_per_contact__call
	interevent_time__text
	interevent_time__call

Indicators over weeks take a single value per week and measure the mean and std over several weeks. Indicators within weeks admit multiple values within each week which are aggregated by min, max, mean, std, skewness, kurtosis. The weekly aggregated values are then aggregated over consecutive weeks using mean and std.