blind test

Our function can be expressed as:

x_act(t)m_share(t_d) = y_ref(t)../f/f_people(t) + y_out(t)

where x_act is the number of activities, m_share is the market share, y_ref reference data - cachier receipts, ../f/f_people people per receipt, y_out people outside the shop, t time, t_d day

reference vs reference

To check the capability of the prediction we take as reference the customer data and as activities the customer data multiplied by a random noise

reference vs reference + gaussian noise, a single location might have low correlation

reference vs reference + gaussian noise, overall sum neutralizes the gaussian noise

score gauss noise 50%

score gauss noise 90% no smoothing, final score on 30 days

score gauss noise 30%, final score on june

score gauss noise 20%, final score on june

day correlation mapping

We take the first 20 cilacs close to a poi and we calculate activites on daily basis.

Activities are processed with a 20km previous distance filter and we match activity chirality with the poi chirality.

6% of the total cilacs correlates over 0.6 with reference data.

The sum of the activities over all country is

no day filtering

We have to filter out bad days

bad day filtering

We perform a weekday correction

weekday correction

2d correlation between cilac patterns

country adjustment

etl - lowcount - wday - filter - reg 00 - 04 - 22 - 30 - 41

iteration of the new mapping scoring

blind test on real data

test on performance on learn & play on same days

smoothing correlates neighboring events and improves the score, june in blind test

curve blind, june

curve blind_single2

curve blind_single

scoring on the different learning steps until blind test

learn play_randomDays

correlation over locations