Skip to content

Add how fast each night is booking as a model input #103

Open

Description

@selvarajuv

Part of Pair 2: booking data into the model (Alex + Heather), under #88. Stretch for this week. Blocked by #88.

Problem

The model only knows the calendar, events and weather for a future night. It doesn't know how that night is actually selling, which is usually the strongest sign of extra demand: if a night is already 80% full a month out when similar nights are at 50%, it's underpriced.

Goal

The model sees how full each night is and how fast it's filling, both when learning from past nights and when pricing future ones.

Scope

This week (training and measurement):

Later (serving):

Considerations

  • A rate depends on when it's set: a night 60 days out and the same night 2 days out need different inputs. Decide which "days ahead" to train on (e.g. one row per night per several lead times, or a fixed lead time).
  • "% sold so far" for future nights uses total rooms, not rooms − ROOS, because future nights' ROOS isn't imported (decision 0037).
  • If the latest upload is old, the inputs are wrong; use the upload date and flag when it's old.

Acceptance criteria

  • The new inputs are computed for training from the rebuild.
  • Results with and without them are on Test which inputs improve the rates #93's backtest.
  • Tests cover the inputs on sample bookings, including cancellations.
  • (Later) Pricing computes them the same way from the latest upload, and recommendations change when a night's bookings pick up.

Metadata

Metadata

Labels

featureIntroduces a new and complete featuremlModel training, feedback in training, features, evaluation

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions