We previously measured that revenue management efforts are able to achieve around 64.23% of the in month revenue potential for the month of April 2024.
There is a big difference in decision count between RandomForest and ATLAS Embedding DNN v2, this is because ATLAS Embedding DNN always get restricted due to price change threshold of 15% or 20%, and below minimum rate price. This caused a lot of revenue to be directed to ATLAS RandomForest just because it gets more chances and executions.
This is also applicable to MARS since price prediction beyond treshold is also not logged in pricing_decision table.
Problem
Root Causes:
If we remove all the restrictions on ATLAS Embedding DNN, we couldn't ensure 100% reliability that the model will always predict price within common sense to the MA ADR, because the MA ADR in the first place is wrong (read below)
A lot of decisions from ATLAS Embedding DNN and MARS does not go through the send stage, causing prices to be outdated.
MA90 ADR per Unit sometimes average bookings between high end unit type and low end unit type together causing the MA90 ADR per unit to be most of the time higher or sometimes lower than common price of the room_id.
Lack of explainability is making it harder to justify the model's predictions
We notice that some property in potential revenue list such as Bong Hostel has a lot of unnoccupied rooms despite alteration of prices.
Solutions
If pricing prediction > treshold, we can adjust final price to be exactly the treshold value, ensuring pricing decisions to be updated and more contribution from respective models.
Move validation of pricing such as minimum rate and out of bound treshold to process stage instead of model/send stage.
Ensure that data logging in pricing_decision table has been completed in the process stage.
Ensure that we are able to log how many pricing decisions ave been pushed through this method.
Update feature collection for MA ADR to accomodate the following
Recognize the seasonality of the NAB dates and get the ADR from the same seasonality group.
SUM the ADR according to the uint type aand quantity to get a better MA ADR value
Reduce cases of gateway timeout that causes inference to be empty and no decision made.
After the number of pricing prediction > treshold as been significantly reduced to <50/day, we can make escalation status = 1 to avoid it being sent, but it requires manual intervention from pricing ops.
Description
We previously measured that revenue management efforts are able to achieve around 64.23% of the in month revenue potential for the month of April 2024.
There is a big difference in decision count between RandomForest and ATLAS Embedding DNN v2, this is because ATLAS Embedding DNN always get restricted due to price change threshold of 15% or 20%, and below minimum rate price. This caused a lot of revenue to be directed to ATLAS RandomForest just because it gets more chances and executions.
This is also applicable to MARS since price prediction beyond treshold is also not logged in pricing_decision table.
Problem
Root Causes:
Solutions
Measurement metrics
Current measurements:
Period of 2024-04-01 - 2024-04-29
SLA
RM:
ATLAS:
MARS:
Updates: