NP-RM-04 · Demand Forecasting & Modelling

📊

Process Flow Diagram (BPMN)

📋

L4 Process Steps

Step	Step Name	Role / Swim Lane	System	Input	Output	KPI	Dec?	Exc?
Phase 1 1.1	Extract historical booking & ticketing data	Revenue Management Analyst	Amadeus Altéa PSS	Scheduled forecast cycle trigger (daily/weekly batch)	Raw booking history file by O&D, cabin, fare class and departure date	Data completeness ≥ 99% of expected PNR records	N	N
1.2	Ingest external demand & traffic signals	Revenue Management Analyst	Amadeus SkyCAST	OAG forward schedule feed, MIDT booking data, tourism board indices	Enriched demand signal dataset merged with internal booking history	External signal ingestion latency ≤ 2 hours from OAG publish	N	N
1.3	Validate & cleanse data for anomalies	Data Quality Analyst	AWS S3 / Redshift	Raw booking history + external signals from steps 1.1–1.2	Cleansed, deduplicated demand dataset with anomaly flags	Data quality score ≥ 95%; outlier rows flagged < 2% of total	Y	Y
1.4	Segment data by O&D, cabin and booking curve	Revenue Management Analyst	Amadeus SkyCAST	Cleansed demand dataset	Demand time-series per O&D × cabin × day-of-week × season band	Segment granularity covers ≥ 95% of revenue-bearing O&D pairs	N	N
Phase 2 2.1	Identify constrained booking observations	Revenue Management Analyst	Amadeus Revenue Management (NRM)	Segmented demand series + historical AU and sell limit records	Flagged constrained observations (departures where AU was binding at close)	Constrained-observation detection rate ≥ 90% vs manual audit sample	N	N
2.2	Apply EM unconstraining to constrained series	Revenue Management Analyst	Amadeus Revenue Management (NRM)	Constrained observation flags + observed booking counts	Unconstrained demand estimate per O&D × fare class × departure date	Unconstraining variance (CV) ≤ 5% on routes with ≥ 90% load factor	Y	Y
2.3	Validate unconstrained demand estimates	Senior Revenue Management Analyst	Amadeus SkyCAST	Unconstrained demand series from step 2.2	Validated unconstrained demand series approved for model input	Unconstrained mean ≤ 1.5× constrained mean on non-peak markets	N	N
Phase 3 3.1	Select forecasting model type per market	Revenue Management Analyst	Amadeus SkyCAST	Validated unconstrained demand series, market volatility scores	Model configuration file: additive vs multiplicative per O&D × cabin	Model selection logic applied to 100% of active O&D pairs within 1 hour of data readiness	N	N
3.2	Apply special-event & seasonality overlays	Revenue Management Analyst	Amadeus SkyCAST	Model config + airline event calendar (holidays, conventions, sports, disruptions)	Seasonality-adjusted demand series with event uplift / suppression factors	Special-event adjustment coverage: ≥ 85% of Top-100 events per quarter captured before forecast run	N	Y
3.3	Run O&D demand forecast in SkyCAST	Revenue Management Analyst	Amadeus SkyCAST	Seasonality-adjusted unconstrained demand series + model config	Point and probabilistic demand forecasts per O&D × cabin × booking horizon	Forecast MAPE ≤ 12% across all O&D pairs; P90 forecast coverage ≥ 80% of actuals	Y	Y
3.4	Disaggregate O&D forecast to fare-class level	Revenue Management Analyst	Amadeus Revenue Management (NRM)	O&D demand forecasts from step 3.3 + fare class demand share curves	Fare-class level demand forecast input for EMSRb optimisation	Fare-class forecast disaggregation completed within 30 min of O&D forecast publish	N	N
Phase 4 4.1	Analyst review of system-generated forecasts	Senior Revenue Management Analyst	Amadeus Revenue Management (NRM)	Fare-class demand forecasts from step 3.4, booking pace actuals	Reviewed forecast report with analyst annotations and override candidates	Analyst review completed within 4 hours of forecast publication; < 15% of O&D pairs require manual override	Y	N
4.2	Apply market-level manual forecast adjustments	Senior Revenue Management Analyst	Amadeus Revenue Management (NRM)	Override candidate list from step 4.1, competitive intelligence, sales intelligence	Adjusted demand forecast with analyst override log	Override accuracy (post-departure MAPE for overridden markets) ≤ 10%	N	N
4.3	Publish approved demand forecast to NRM	Revenue Management Analyst	Amadeus Revenue Management (NRM)	Approved fare-class demand forecast (system + analyst overrides)	Active demand forecast loaded into NRM for EMSRb optimisation cycle	Forecast publish latency ≤ 30 min from analyst sign-off; 100% of departures within 330-day booking window covered	N	Y
Phase 5 5.1	Track post-departure forecast accuracy	Revenue Management Analyst	AWS S3 / Redshift	Departed flight actuals (load factor, revenue, fare-class mix) vs forecast	Forecast accuracy scorecard: MAPE, bias, coverage by O&D, cabin, horizon	Rolling 30-day network MAPE ≤ 12%; bias (mean signed error) within ±2%	N	N
5.2	Detect systematic bias or model drift	Senior Revenue Management Analyst	Tableau / Power BI	Forecast accuracy scorecard from step 5.1	Bias / drift diagnostic report with affected O&D segments identified	Bias alert triggered when mean signed error exceeds ±3% for ≥ 5 consecutive departures on same O&D	Y	N
5.3	Trigger model recalibration & parameter update	Revenue Management Analyst	Amadeus SkyCAST	Bias / drift diagnostic report + recalibration threshold breach flags	Updated model coefficients and booking-curve parameters deployed to SkyCAST	Recalibration cycle completed within 48 hours of bias-alert trigger; post-recalibration MAPE improvement ≥ 2 percentage points	N	Y

📋

Process Attributes

Identification

Process IDNP-RM-04

L1 DomainNetwork Planning & Scheduling

L2 ProcessRevenue Management

L3 NameDemand Forecasting & Modelling

L4 Steps17 across 5 phases

Decision Gates5 (all with iteration loops)

Exceptions6 documented

Swim Lanes (Roles)

Revenue Management Analyst

Data Quality Analyst

Senior Revenue Management Analyst

Systems & Tools

Amadeus Altéa PSSAmadeus SkyCASTAWS S3 / RedshiftAmadeus Revenue Management (NRM)Tableau / Power BI

Key Performance Indicators

Extract historical booking & ticketing dataData completeness ≥ 99% of expected PNR records

Ingest external demand & traffic signalsExternal signal ingestion latency ≤ 2 hours from OAG publish

Validate & cleanse data for anomaliesData quality score ≥ 95%; outlier rows flagged < 2% of total

Segment data by O&D, cabin and booking curveSegment granularity covers ≥ 95% of revenue-bearing O&D pairs

Identify constrained booking observationsConstrained-observation detection rate ≥ 90% vs manual audit sample

Apply EM unconstraining to constrained seriesUnconstraining variance (CV) ≤ 5% on routes with ≥ 90% load factor

Validate unconstrained demand estimatesUnconstrained mean ≤ 1.5× constrained mean on non-peak markets

Select forecasting model type per marketModel selection logic applied to 100% of active O&D pairs within 1 hour of data readiness

Airline-Specific Risks & Pain Points

Altéa PSS booking history exports can lag 4–6 hours in peak windows, misaligning the daily forecast cycle and causing stale AU recommendations to persist in NRM overnight

MIDT data covers only GDS-booked itineraries, creating blind spots for NDC and direct-channel bookings — up to 30% of volume on some LCCs — leading to systematic under-forecasting on digitally distributed routes

COVID-era booking collapse and recovery create structural breaks in training data; models trained on pre-2020 history without explicit break-point adjustment produce load-factor forecasts biased 6–12 pts below actuals on leisure routes

Thin O&D markets (< 200 observations per departure) produce statistically unstable booking curves; SkyCAST defaults to network-level priors which can over-smooth volatile new-route ramp-ups

NRM does not retain intra-day AU change history at sufficient granularity; when AUs were changed multiple times within a day, the system only logs end-of-day state, causing the unconstraining algorithm to miss intermediate constraints

Expectation-Maximisation unconstraining can diverge on very high-load routes (LF > 95%) where almost every observation is constrained — NRM falls back to simple censored-mean imputation which understates true demand by 8–15%

Inputs / Outputs

Primary InputScheduled forecast cycle trigger (daily/weekly batch)

Primary OutputUpdated model coefficients and booking-curve parameters deployed to SkyCAST

←PreviousNP-RM-03 · Overbooking Policy Management NextNP-RM-05 · Group Sales Revenue Management→