← All posts

R2CCP Bid Prediction

Multimodal R2CCP forecasting · 8 context models · 500K Monte Carlo decisions

Built with Python · R2CCP · Conformal prediction · Entropy regularization · Monte Carlo simulation

Synthetic bimodal distribution comparing a collapsed cumulative interval with per-bin conformal regions that preserve the low-density gap between modes
Synthetic illustration of the interval-collapse fix

WHY

PQ bids are evaluated by both technical and price scores. As the technical score falls, viable price bands split around invalid regions, so actual bid behavior becomes multimodal rather than forming one central peak. A useful decision system had to preserve those separated regions and estimate the win probability of each candidate rate instead of returning one point or one merged interval.

Modeling dataset 69,934 records

Chronological train and validation split

Validation 13,984 samples

90.73% weighted conformal coverage

Operational result +35% KPI

Group-company PQ bid performance

HOW

Preserve the distribution, then simulate the decision

Context-specific calibration keeps distinct modes separate before simulation evaluates candidate actions.

  1. 01 Context routing

    Combine ranking position Q1/Q2 with four bid-rate-difference quartiles.

  2. 02 R2CCP distribution

    Train eight entropy-regularized context models on a chronological split.

  3. 03 Per-bin conformal set

    Retain only bins above the calibrated threshold and preserve disjoint intervals.

  4. 04 Monte Carlo decision

    Sample competing outcomes 500,000 times and estimate the win probability of each candidate rate.

RESULT

Eight-context chronological validation

ContextCoverageAverage interval length
Q1-BRD191.39%0.0149
Q1-BRD292.23%0.0194
Q1-BRD388.07%0.0166
Q1-BRD491.21%0.0169
Q2-BRD186.14%0.0220
Q2-BRD281.41%0.0208
Q2-BRD393.92%0.0285
Q2-BRD494.77%0.0534

Weighted coverage: 12,688 / 13,984 = 90.73%.

Coverage chart for eight Q by BRD context models over 13,984 chronological validation samples with 90.73 percent weighted coverage
Chronological validation · 13,984 samples · weighted coverage 90.73%

Contribution

  • Diagnosed multimodal interval collapse and rebuilt inference with per-bin conformal thresholds and entropy regularization.
  • Designed the eight-context chronological evaluation and connected distribution forecasts to 500,000-iteration candidate simulation used in bid decisions.

Evidence

Source code R2CCP modeling, calibration, context routing, and simulation