A Coggle Diagram about Model Variables
, Block Perspective Assessability
, Online Check-In
, Clustering salons
, Salon Network
More detailed ranking of available Great Clips opportunities and how the new salon would change that ranking
, Block level aggregation -
Split Blocks if > x% of CCs?
Find Max % and evaluate
When rolled up annually
% of NBGG at Block level
, Salon Age
Pay careful attention to the age of the impacted salon. Particularly if the salon has been open less than 2 years.
, Global Customer
What proportion of sales in the Block come from customers who are NBGorG at another salon
How comfortable is the Block with moving between salons
Do they stay put because of limited options or because of loyalty?
, Discount Sensitivity
If a greater proportion of sales from a block are on discount do they transfer more (Seeking new discounts) or less (Still getting discounts at existing)
and Regional Draw - Salon Level Variable
How large is the trade area of the existing salon
Is it a neighborhood location or a regional center
What proportion of customers are nearer to other salons
), Third Party Data
, Kavita (D&B?)
, Agg Data
Every 2 years - Intermittent
Plugs into CRM but static file as the source
DUE FOR AN UPDATE DONE
US and Canada
and Igeocode - Canadian Lat/Longs
Incremental gain over MelissaData match
Hard match on Postal code look up to Lat/Long
Duplicate Canadian spotting workflow
Updated every 2 years? inconsistantly
), Canada - In scope?
(Recommend Probabilistic approach
x% chance of being above or below 10%?
, Variables that were not predictive or intentionally omitted
, Present to RE committee
and Weighting of variables?
Credit Score example
), 50% of sales
(Only allow salons with > 50% of known sales mapped into sample set
Customer Counts Not Sales
Change in Production
, Continue to suppress > 50% salons from analysis
and Move calculation to ETL process instead of run-time
(Oversample at the Block level
Activity in many block groups has nothing to do with the new salon opening. Measuring changes in these blocks obscures the nature of the relationships that do matter. How do we limit analysis to only block groups that can tell us something about transfer
, Examine actual impacts at the block group level In relation to other variables. Break by cluster. High, Medium, and Low impacted Blocks
and Oversample at salon level
Most impacts are < 10%. Limit the number of these salons in the analysis. Oversample the mid and high impact scenarios.?There are probably still learnings at the block level in these scenarios even if we don't model at the salon level.We may get into sample size issues if we break by 5 density classes.
(Add New, Better, Good Great Count
, OCI count, OCI PASS, Wait Time, Request Data
, Global Customer Activity
, % of known sales - CCs instead now? Both?
and Distance from Block to salon - Which aggregation? # KEEP CURRENT CALCULATION
) and Database creation
), Special Evaluation
What role do they play, what impact would there be if we removed them
Also a sensitive topic. Understand the contribution before communicating
), Sample Data
(Attach Scott's Global data to salons
How many years back?
How many to exclude for Validation?
How many salons will become eligible in the rest of the year. Vs. How many do we want to evaluate?
(Visibility Into variables
), Block Customer
(Backfill with one year of data
Put into production for evaluation set
Final structure to Cuong
), Time Series - Can we predict how much impact you will have along the way in addition to a final number? Quarters post-open?
Can we measure progress against that and intervene?
, Globalview dropouts
Needed to see visits from
WAITING ON SABG
One time forensic analysis but capture global in production
, Supplement Missing Data
Salon under 50 % don't receive a PSA but there is a varying degree between eligible salons. Estimates are based off of raw customers at the salon level. If a salon is low on sales from a block that doesn't mean they aren't there. We shouldn't be predicting growth
, Estimate change in CC not Sales
and Average distance traveled more predictive than distance from the average point? What are the calculation implications? - NOT DOING THIS. Individual calculations at run-time too high if not during ETL - New salon distances