IE437 · Assignment 1 Case Study

Cinema Programming
Response Model

Combining audience response prediction with dynamic scheduling
A next-generation cinema scheduling system

Predict→ Respond→ Optimize→ Learn
IE437Assignment 1Case scope
→ / Space Next · ← Previous · F Full screen · P Print
02Executive Summary

Turn the baseline forecast into
better film programming decisions

Build on the cinema chain's Base Demand Forecast, learn how audiences respond to scheduling actions, optimize feasible schedules, and continuously refine future schedules using actual bookings and attendance. Together, they form the proposed AI Programming Intelligence for Cinema.

01

Maintain and improve Base ForecastPotential Demand

Predict a film's potential demand while retaining and improving the cinema chain's existing forecasting assets.

02

Add Audience ResponseAction → Response

Predict how screen, showtime, frequency, and format decisions change audience choices.

03

Programming OptimizationDecision Engine

Find the best feasible schedule while respecting operational constraints.

04

Dynamic Re-programmingRolling Horizon

Continuously optimize adjustable future schedules using actual bookings and audience response.

From Predict → Allocate to Predict → Respond → Optimize → Learn
IE437 · Cinema PROGRAMMING CASE
03Part I — Baseline operating structure

Forecast demand and build each site’s timetable,
then connect the two through response

Prediction — occupancy forecasting Model ② · Occupancy forecasting
P1Forecast simulation · National market forecastNATIONAL
↓
P2Company-wide allocation — screen operating planCOMPANY
↓
P3Site-level allocation · FeedbackSITE
↓
P4Weights derived from site performance dataWEIGHT
Programming — automated scheduling Model ① · Automated scheduling
A1Scheduling criteria — operating and cleaning hoursSETUP
↓
A2Apply automated scheduling rulesRULE
↓
A3Refine rules for site characteristicsSITE-FIT
↓
A4Produce site-level screening schedulesSCHEDULE
This proposal builds on that architecture. The two domains connect more effectively through a stronger decision layer.
Operating structure · Demand forecasts → screen allocation → site timetables
04Part I — Baseline operating structure

The baseline approach is sound:
forecast demand, then allocate it

FILM A — EXPECTED NATIONAL AUDIENCE 5.0M Site A Site B Site C Site D Allocate shares Allocate shares Allocate shares Allocate shares Allocate screens and showtimes
THE CONVENTIONAL LOGIC PREDICT → ALLOCATE

Base Demand Forecast is essential.
Every schedule starts with a demand estimate. This forecast is a core Cinema asset.

IllustrativeConceptual allocation diagram — illustrative values
05Part I — Existing cinema operations

One important assumption
is embedded in this approach

“Audience demand is
already fixed before scheduling.”

Under this assumption, scheduling assigns predetermined demand to screens and times as an allocation task rather than a response problem.

FIXED DEMAND 🔒Demand = 5.0M
ALLOCATE
Is demand really fixed independently of schedulingin practice?
← This question is the starting point of our proposal
06Part II — A different perspective

Scheduling changes audience choices
as well as allocating demand

Potential Demand + Site Context + Programming Action → Realized Audience
A

Film A — scheduling scenario simulation

AUDIENCE RESPONSE SIMULATOR

Screen sizeShowtimeFrequency FormatCompeting titles

Adjust three actions: time, screen size, and frequency. The Response Model also predicts format and competing-film effects.

Showtime
Screen size
Daily frequency
Film A AttendanceFilm A attendance
1,050
Other-title CannibalizationOther-film cannibalization
−340
Total Site AttendanceTotal site attendance
4,110
Baseline ┊ = 3,400 site visitors without Film A (synthetic). Film A's potential demand is identical in every scenario.

Increasing frequency from 2 to 5 shows does not increase attendance proportionally; cannibalization rises faster. A good schedule must account for both diminishing returns and cannibalization.

Scheduling actions themselves shape demand.
Illustrative ExampleAll figures are illustrative synthetic values
07Part II — A different perspective

Even an accurate forecast of 5 million viewers
leaves the cinema chain’s captured audience open to change

National Potential Audience · Fixed 🔒 5.0Mviewers (national potential audience)

This figure stays the same under either schedule A or B.
It represents the film's underlying potential demand.

3 prime-time shows Large-screen allocation Balanced showtime spacing Premium formats
Realized cinema attendance 1.20M= 24% of potential
B 0.90M
SCALE = POTENTIAL 5.0M
Prediction Accuracy≠Programming Quality With the same forecast accuracy, scheduling quality changes the audience Cinema actually captures.
Illustrative1.20M / 0.90M are synthetic illustrations. Focus on the A–B difference, rather than the absolute share.
08Part II — A different perspective

Distinguish the film’s potential demand
from the cinema chain’s realized demand

Potential Demand — the film's audience potential
  • Underlying box office appeal
  • National · Regional potential demand
  • Release week · Awareness · Viewing intent
ESTIMATED BY — BASE DEMAND FORECAST
Programming
converts
Scheduling converts potential
into realized demand
Realized demand — audience captured
  • Screen Screen
  • Showtime Showtime
  • Frequency Frequency
  • Format Format · Competing-film placement Competing titles
ESTIMATED BY — AUDIENCE RESPONSE MODEL
Retain the existing Base Forecast as the Response Model's core context for every decision.
POTENTIAL → PROGRAMMING → REALIZED
09Part III — System architecture

Prediction estimates outcomes;
Programming chooses actions

PREDICTION
“How will audiences respond to this schedule?”
Submodule 1Base Demand Forecast

Estimate potential demand by film, site, and date — retain and improve the cinema chain's assets

Submodule 2Audience Response Model

Predict attendance given a scheduling action — the new response layer

OUTPUT
Expected attendanceSeat occupancy Revenue proxyUncertainty (confidence interval)
PROGRAMMING
“Which schedule should we choose?”
Decision EngineProgramming Optimization

Generate feasible schedules, evaluate them through Prediction, and search for the best option

INPUT
Repeated Prediction Service callsOperational constraints
OUTPUT — executable schedule
Movie × Screen × Showtime × Format
⇄
Prediction estimates outcomes; Programming chooses actions.
ANCHOR — this distinction guides the rest of the presentation
10Part III — Prediction ① Base Demand Forecast

Base Forecast remains essential:
it estimates the film’s potential demand

Film attributes (genre · director · cast) Awareness · Viewing intent Reviews · Buzz Release week Advance-booking schedule Regional · Site characteristics
Potential Demand Curve — estimated audience potential
Release week D-14 D+28 PEAK POTENTIAL Demand curve — conceptual example
OUTPUT — Potential Demand by Film × Site × Day
Retain the cinema chain's existing forecasting assets and benchmark and refine them as the Response Model's core context.
PREDICTION ① — maintain and improve existing assets
11Part III — Prediction ② Audience Response Model

The next question is:
“How will audiences respond to this schedule?”

Attendance  =  f ( Base Demand ,  Context ,  Programming Action ) BASE DEMAND FORECAST Film demand potential SITE CONTEXT Site context · Bookings AUDIENCE RESPONSE MODEL Attendance for this schedule E [ Attendance | Action ] EXPECTED REALIZED ATTENDANCE Attendance + uncertainty PROGRAMMING ACTION — Screen size Showtime Number of shows Format Competing films
Diminishing ReturnDiminishing returns from added shows
SubstitutionAudience shifts across the same film's shows
CannibalizationCannibalization across films
Format UpliftDemand uplift from premium formats
PREDICTION ② — the new response layer · Four representative effects to predict
12Part III — Requirements for response learning

“Larger auditorium → more viewers” can lead to
a false conclusion about scheduling effects

Larger auditorium LARGER AUDITORIUM Higher attendance ATTENDANCE ↑ CORRELATION Expected popularity HIDDEN COMMON CAUSE Films expected to perform well receive better schedules Films expected to perform well and already attract more viewers Correlation ≠ the causal effect of scheduling
Did the larger auditorium bring more viewers,
or did a promising film
receive a larger auditorium?

Historical schedules were chosen based on expected popularity(endogeneity). Validate scheduling effects through Historical Replay and careful identification, rather than interpreting correlation as causation.

▸ Press → to reveal: ① correlation ② hidden path ③ conclusion
Addressing this trap is essential to Response Model design
13Part IV — Programming Engine

Compare thousands of feasible schedules
to find the best option

Show allocation heatmap — site × film (cell: shows · expected attendance) RULE-based initial allocation
Site A Screening schedule — click a site name to inspect its schedule
10:0013:0016:00 19:0022:0024:40
Chain-wide performance — Prediction ↔ Programming
Reference schedule — no site-specific response—
Response-aware reoptimization: Best —
0Prediction calls (API)
0Search iterations
0Constraint violations
Changes — ITERATION · action · Δ
    Chain constraints: film show caps · ≥2 sites per film · ≤5 shows per film/site · format fit (C: ≥180 seats; E/F: no large screens) · 15-min cleaning · operating hours
    Prediction supports more than a single evaluation — Programming uses repeated calls to search the decision space.
    Synthetic Instance6 sites with different audience and screen profiles — live in-browser local search · amber cells = changed allocations
    14Part IV — Programming Engine

    Keep the cinema chain’s operating rules,
    and express them as optimization constraints

    Operating rules — reference process
    Operating hoursOPERATING HOURS Cleaning timeCLEANING Screen quotaSCREEN QUOTA Premium-screen share capPREMIUM CAP Format compatibilityFORMAT FIT Protect existing bookingsBOOKING LOCK Operating policies · ExceptionsFIELD POLICY
    STRUCTURED
    AS
    Inside the Programming Optimizer
    Hard Constraint MUST SATISFY

    Mandatory: operating hours, cleaning time, screen quota, and booking protection. Every recommended schedule must pass feasibility checks.

    Soft Policy PENALIZED

    Preferred: show spacing, screen assignment practices, and premium-screen policies. Violations receive penalties.

    Objective MAXIMIZED

    Optimize attendance, occupancy, and contribution value. Define the objective jointly with Cinema.

    Operating rules guide the AI — they become the operating grammar it must respect.
    Scheduling constraints · Operating hours, turnaround, capacity, format and booking rules
    15Part IV — Target Architecture

    The proposed film programming system:
    the complete architecture

    Film / Market / Booking Data PREDICTION Base Demand Forecast Potential demand · Existing forecast Audience Response Model Scheduling response · New model PROGRAMMING Programming Optimization Generate · Evaluate · Select Candidate scheduling action Predicted response · Score Cinema rules / constraints Hours · Cleaning · Quota · Freeze Recommended Schedule Observed bookings / attendance LEARNING LOOP Retrain from actuals DYNAMIC RE-PROGRAMMING Bookings → optimize unopened shows
    PREDICTION (blue · teal) and PROGRAMMING (purple) are designed as separate services
    16Part V — Dynamic Re-programming

    Once the initial schedule is set,
    better information starts arriving

    Initial forecastINITIAL FORECAST
    ↓
    Initial schedule setINITIAL SCHEDULE
    ↓
    Bookings openBOOKING OPEN

    After this point, the schedule is often treated as a fixed deliverable 🔒 while new signals continue to accumulate.

    Signals arriving after bookings open
    Booking velocityBOOKING VELOCITY
    Sales by time slotTIME-SLOT SELL-THROUGH
    Response differences by siteSITE-LEVEL RESPONSE
    Changes in audience mixAUDIENCE MIX
    Competing schedules · ResponseCOMPETITOR RESPONSE
    Reviews · BuzzREVIEWS / BUZZ
    Opening-day performanceFIRST-DAY ACTUALS
    Should the schedule stay unchangedas signals improve?
    PART V — Dynamic Re-programming answers this question
    17Part V — Dynamic Re-programming

    The next step is AI that
    operates and improves schedules

    Actual Response
    Booking Index
    (Forecast = 100)     100On forecast
    FORECAST 100 100 128 145 MON WED FRI

    The initial schedule is set and bookings are open.

    Adjustable schedule — Sunday SCHEDULE v1
    Screen 1 — 300 seats LARGE
    Film A10:30
    Film A13:30
    Film B16:30
    Film A19:30
    Film C22:00
    Screen 4 — 120 seats SMALL
    Film B11:00
    Film C14:00
    Film B17:00
    Film B20:00
    10:0014:0018:0022:00
    🔒 Bookings open — locked Reoptimized change Show removed
    Initial Programming→ Actual Response→ Update→ Re-programming ↻ ROLLING — continuous reoptimization from actual response
    Illustrative · SyntheticShows already open for booking remain unchanged — Freeze rules on the next slide
    18Part V — Dynamic Re-programming

    Reoptimize while preserving
    customer commitments and operational stability

    NOW
    🔒FROZENBookings already open — locked by default
    LIMITEDLimited changes — operator approval required
    OPTIMIZABLEActive reoptimization — before bookings open
    0h+24h+48h+72h →
    4-week rolling schedule — actuals → Freeze rules → unopened-show optimization → learning DAY 0 · Initial schedule
    Cumulative attendance advantage — rolling vs. fixed initial schedule
    +0.0%Cumulative +0 viewers
    Weekly forecast error — learning
    Schedule changes
    0Cumulative changes · ≤4 per update · weekly retraining
    ActualsFrozenLimitedRescheduled
    Refine eligible windows through Rolling Horizon Optimization with operational stability.
    Synthetic rolling illustration · 0–24h / 24–48h / 48h+ are example windows; operating policies define the final Freeze rules.
    19Part VI — Validation plan

    Validate four hypotheses
    in sequence

    H1Base Forecast — potential demand

    Can we estimate underlying film and site demand reliably?

    H2Action Response — audience behavior

    At equal Base Demand, can we explain response differences by screen, time, frequency, and format?

    H3Optimization — scheduling decisions

    Does response-aware programming improve business KPIs over current rules and schedules?

    H4Dynamic — operational reoptimization

    Does reprogramming from actual response improve on a fixed initial schedule?

    Validation builds from H1 to H4 — each stage depends on the preceding stage
    20Part VI — Proposal scope

    Build one closed loop
    from forecasting to validated scheduling decisions

    D1

    Base Forecast benchmark · Improved model

    Diagnose and improve performance against existing forecasts

    D2

    Audience Response Model

    Predict audience response to scheduling actions

    D3

    Programming Optimization Engine

    Constraint-based scheduling decision engine

    D4

    Historical Replay · Counterfactual Simulator

    Test “what if we scheduled differently?” using historical data

    D5

    Dynamic Re-programming recommendation module

    Rolling Horizon reoptimization recommendations

    D6

    Validation Report · Operating guidelines

    Hypothesis results and documented operating criteria

    D7

    API · Integration specification

    Integration Spec for EAI / DX AI connectivity

    CLOSED-LOOP

    D1–D7 form one integrated loop:
    Predict → Respond → Optimize → Learn

    Outside the scope Full replacement of existing theater business systems A new company-wide UI platform
    Add an Intelligence Module to the cinema chain's existing system assets
    21Part VI — Data for the decision loop

    Connect film, booking and site data
    to each part of the scheduling decision

    Base DemandEstimate potential demand
    Film attributesFilm Viewing intentBooking AwarenessBooking Reviews · Buzz (quantitative · qualitative)Booking Advance-booking scheduleBooking Release date · First screening dateFilm Genre · Director · Cast · SynopsisFilm
    ResponseLearn audience response
    Show countSite Scheduled seatsSite Screening date · Booking dateBooking Booking opening timestampBooking Booked seats · TicketsBooking Screening formatFilm Aggregated audience segmentsBooking Sales channel · Film occupancyBooking
    Constraint · ProgrammingScheduling constraints
    Operating hoursSite Cleaning timeSite Premium-screen share capSite Screen-quota informationSite Premium-screen space · TechnologySite Screen name · Operating floorSite
    Decision inputs · Demand context → schedule and audience response → feasible timetables
    22Part VI — The most important additional data

    Response learning needs
    “How did the schedule change?”

    OPEN PLAY D-7D-5 D-3D-2 D-0 Schedule v1Schedule v2 — change Δ Bookings openSchedule adjustment Δ Actual screening

    A final schedule shows only what was screened. Versioned snapshots and change history align schedule changes with later booking and audience responses — the key raw material for response learning.

    Required Fields
    • Schedule snapshot timestamp
    • movie / screen / showtime / seats
    • Revision history
    • Booking state at revision
    • Change reason · Operator override (if available)
    Action History helps identify how responses vary with scheduling actions and support stronger effect estimation.
    Key data that determines H2 (Action Response) validation depth
    23Part VI — Data-to-decision interfaces

    Turn operating inputs into
    a usable scheduling recommendation

    Operating inputProcessing stepDecision componentOutput for the operator
    Film attributes · Release date · Booking signalsEstimate film demand by planning periodBase ForecastDemand outlook for each film
    Schedule versions · Booking and attendance outcomesRelate film, site and time choices to audience responseResponse ModelExpected audience for a candidate timetable
    Screen capacities · Formats · Operating hoursDescribe the screens and time available for each siteScreen inventorySite-specific planning workspace
    Runtime · Cleaning time · Quotas · Booking locksCheck hard rules and apply scheduling policiesConstraint EngineFeasible schedules and reasons for rejected moves
    Reference schedule · Candidate scoresCompare film allocation, showtimes and expected attendanceProgramming EngineRecommended schedule with changes highlighted
    Selected schedule · Operator edits · Observed outcomesPublish the timetable and feed outcomes into the next cycleExecution and feedbackBooking-ready timetable and updated decision history
    Operating interfaces · Source inputs → decision engines → timetable and feedback
    24Part VI — Three-stage implementation plan

    Build in three stages:
    feasible timetable → response model → integrated planning

    IMPLEMENTATION FLOW—Feasible schedules→Better decisions
    PRE-TRACK Load the reference timetable, film and screen data, operating rules and planning objective Shared inputs for all three stages
    STAGE 1Feasible scheduling · Model ①
    Operator track
    Create a timetable from screen and film inputs
    Review showtimes and operating rules
    Modeling and search track
    Pack screenings within opening and closing hours
    Check runtime, turnaround and film requirements
    Keep the reference timetable for comparison
    STAGE 2Response-aware decisions · Models ① + ②
    Operator track
    Compare film allocations and expected audiences
    Inspect changes for each site and screen
    Modeling and search track
    Estimate demand by film, site and time
    Score reference and candidate timetables
    Search for a better response-aware schedule
    STAGE 3Integrated planning · Models ① + ②
    Operator track
    Review, adjust and publish the selected timetable
    Update the plan as booking signals arrive
    Modeling and search track
    Integrate Base + Response + Programming
    Return audience outcomes to the learning loop
    Keep booked screenings fixed during updates
    Reoptimize the remaining available screenings
    Keep one shared baseline and add capabilities step by step.
    Implementation sequence · Build a feasible timetable → add response-aware search → connect daily operations
    25Part VI — Success metrics

    Measure Prediction and Programming success
    separately

    Prediction KPI — forecast quality
    • MAE / WMAPE Or agreed accuracy metric
    • Ranking accuracy Ranking Accuracy
    • Calibration · Prediction intervals Uncertainty reliability
    • Segment · Site stability
    Programming KPI — schedule quality
    • Total attendance
    • Seat occupancy
    • Revenue · Contribution proxy If available
    • Prime-time · Premium-screen utilization
    • Constraint violations = 0 Mandatory
    Operational KPI — operating quality
    • Scheduling lead time Programming Lead Time
    • Manual adjustment rate
    • Recommendation adoption Adoption Rate
    • Exception · Override rate
    • Lead time to bookings opening
    Define forecast, schedule and operating targets against a shared measured baseline.
    26Part VII — Delivery and operation

    Unite distinct expertise in
    one Decision Modeling framework

    Project team — delivery roles
    • Project / Decision Modeling LeadLead
    • Audience Response / ML LeadPREDICTION
    • Optimization LeadPROGRAMMING
    • Simulation / Integration EngineerPLATFORM
    • Technical Advisor / PIPART-TIME
    COUNTERPART
    INTERFACE
    Role-based
    1:1 collaboration
    Cinema — collaboration team
    • Strategy / Project OwnerDecision
    • Content · Programming domain ownerDOMAIN
    • Theater sales · Operations ownerOPERATION
    • Data / DX AIDATA
    • IT · EAI · Theater business systemsSYSTEM
    1 cinema Decision Owner + 1 Technical Lead — all decisions and issues follow a single escalation path.
    Finalize the detailed responsibility matrix at kickoff
    27Part VII — Delivery and operation

    Prediction and Programming
    also operate differently

    Prediction Lifecycle — Data-driven
    Data Refresh Retrain Validate Deploy Drift monitoring CONTINUOUS
    Programming Lifecycle — Policy · Optimization
    Update policy Validate solver Release KPI review GOVERNED
    DECISION / OUTCOME LOG Record scheduling decisions and actual outcomes to connect both lifecycles — the basis for continuous learning.
    Prediction = Data-driven Lifecycle · Programming = Policy / Optimization Lifecycle
    28Part VII — System integration

    Connect an Intelligence Module
    to a cinema operating architecture

    BASELINE — illustrative cinema architecture Theater business system THEATER BUSINESS · Existing UI EAI DX AI server PREDICTION · OPTIMIZATION DB Store · Display in UI SCREEN / REVIEW INTELLIGENCE MODULE — new · Within/alongside DX AI Base Forecast Service Audience Response Service Programming Optimization Service All calls and results use the existing EAI · DB paths
    1 Theater system requests a schedule 2 Optimizer generates candidates 3 Response Service evaluates candidates 4 Return the best executable schedule 5 Store through EAI in existing DB · Display in UI
    Retain the cinema chain's UI and business systems and improve the decision engine.
    Integration design · Theater business → EAI → prediction and optimization services → DB
    29Part VII — Connecting the operating workflow

    Connect the recommendation to
    the cinema’s daily planning workflow

    Step 1 — Planning workspace
    Prepare the planning runOne objective, one reference and one search loop
    • Load film, booking and screen inputs
    • Set operating rules and the planning objective
    • Generate a response-aware candidate timetable
    →
    Step 2 — Operator review and publication
    Connect the timetableReview changes · Publish schedule Send the selected schedule to booking systems
    • Compare the reference and recommended time slots
    • Accept the recommendation or retain operator edits
    • Publish showtimes through the existing interfaces
    →
    Step 3 — Booking feedback and updates
    Repeat the decision loopUpdate the next planning run
    • Preserve booked screenings — change eligible slots
    • Feed booking and attendance outcomes to Prediction
    • Reoptimize available slots when demand changes
    Carry the decision into operation: prepare inputs → review and publish → learn and replan.
    Daily operation · A recommendation becomes a published timetable and the next planning input
    30Part VII — What the operator gains

    The value goes beyond
    time saved on scheduling

    Business
    • Audience coverageExpected viewers reached by the available screenings
    • Occupancy and revenueTrack audience-to-seat ratio and ticket revenue
    • Premium-screen utilizationMatch film demand to screen capacity and format
    Operation
    • Scheduling timeTime from planning inputs to a publishable timetable
    • Operator review effortRule exceptions, manual edits and review time
    • Replanning speed and stabilityResponse time and number of affected screenings
    Intelligence Asset
    • Action–Response data assetAccumulate decisions and responses
    • Reusable constraint frameworkDocument operational grammar
    • Continuously improving capabilityCompounding value from the learning loop
    Operating metrics · Audience coverage, screen use, planning effort and schedule stability
    31Part VII — A daily operator run

    A natural operating sequence:
    inputs → settings → search → timetable

    The operator starts from a reference schedule, reviews a recommendation, and carries the selected plan into the next cycle.

    01 · Import planning inputs

    Load films, screens, operating rules and the reference timetable into one planning workspace.

    02 · Set the planning goal

    Choose the audience objective, available operating window and search budget.

    03 · Run feasible search

    Evaluate candidate film allocations and showtimes while respecting screen and timing constraints.

    04 · Inspect the recommendation

    Compare expected audiences, screen utilization and changed time slots with the reference.

    05 · Review and export

    Keep the selected plan, record operator edits and export the timetable for publication.

    06 · Learn and replan

    Return booking and attendance outcomes to the models and update eligible future screenings.

    Operator workflow · Import → configure → optimize → inspect → publish → update
    32Part VIII — Bayesian optimization for schedule search

    Use Bayesian optimization to choose
    the next schedule worth evaluating

    Response estimate→ Uncertainty→ Next candidate→ Evaluate→ Update

    Feasible schedule space

    Define film, screen and time decisions under operating constraints.

    Response evaluator

    Score a candidate schedule with a formula, simulator or learned model.

    Surrogate and posterior

    Learn schedule-score estimates and uncertainty from evaluated candidates.

    Acquisition rule

    Balance promising improvements with informative evaluations.

    Candidate evaluation

    Check feasibility, call the evaluator and record the resulting score.

    Update and repeat

    Add the new evaluation; update the posterior until the budget is spent.

    Match the search strategy to the evaluation cost Cheap response calls: local search · Expensive response calls: consider constrained BO
    BO search loop · Feasible candidate → response evaluation → posterior update → next candidate
    33Part VIII — Reading the recommendation

    Compare the reference and recommended plan
    at the level where the operator acts

    Reference Starting timetableBefore search Recommendation Best feasible candidateAfter search Difference Audience and allocationWhat changed Selected plan Operator decisionTimetable export
    Audience

    Expected attendance

    Response score
    Read reference and candidate totalsInspect each film and site contributionIdentify where the audience gain comes from
    Allocation

    Film × site coverage

    Allocation map
    Compare screenings and seats by filmInspect the film mix at each siteConnect allocation changes to demand
    Timetable

    Screen and time changes

    Before and after
    Select a site and screenInspect films, start times and turnaroundReview moves around peak viewing hours
    Feasibility

    Operating rule checks

    Executable timetable
    Read opening, closing and cleaning checksInspect capacity, coverage and film capsResolve any rejected operator edit
    Execution

    Selected schedule

    Timetable output
    Accept the plan or retain the referenceExport showtimes and the decision recordUse the selected plan in the next cycle
    The result is a timetable the operator can understand, review and put into operation.
    Result review · Audience score → film allocation → screen timetable → selected schedule
    34Implementation roles

    Functional roles connect models
    to operating decisions

    01
    Problem framing

    Decision lead

    Define goals, decision variables, validation gates, and the operating objective.

    02
    Forecast and response

    Response modeling

    Estimate potential demand, action response, uncertainty, and calibration.

    03
    Feasible schedules

    Optimization

    Structure constraints, evaluate candidates, and search for useful improvements.

    04
    Operational delivery

    Platform integration

    Connect prediction and scheduling services through APIs and existing systems.

    05
    Domain review

    Operations

    Capture operating rules, review recommendations, and monitor actual outcomes.

    Shared responsibility: problem formulation · validation · operational integration
    IE437 · ASSIGNMENT 1 · FUNCTIONAL RESPONSIBILITIES
    35Next steps — decisions requested

    Turn the proposal into action
    with three decisions

    ASK 01

    Approve a focused prototypePhase 1 — Build & Validate

    Improve Base Forecast; build Audience Response Model and Programming Engine; compare it with the reference schedule. Deliver all D1–D7 outputs.

    First result one feasible improvementCommon inputs · Explicit objective · Reproducible comparison
    ASK 02

    Connect the planning inputsData for the decision loop

    Connect film, screen and outcome inputs; include scheduling Action History and operating rules so demand, response and feasible search use one planning context.

    ASK 03

    Appoint the collaboration structureSingle Escalation Path

    One cinema Decision Owner plus operator interviews and review windows. Use one decision path to resolve policy conflicts and approve schedule changes.

    Prototype and operating scope
    Prototype — model, search and comparison Build a working decision loop: response evaluator, feasible search, reference comparison and operator review; document the data and API contract.
    Operating rollout — integration and monitoring After a useful prototype, connect the recommendation service to the operator workflow and add monitoring, rollback and model updates.
    Decision today→ Inputs · Reference schedule · Rules→ Build · Compare · Review
    Next-step plan · One owner, one reference schedule and one measurable operating objective
    36Closing

    From forecasting 5 million viewers
    to capturing more of that demand at Cinema

    5.0M POTENTIAL DEMAND PROGRAMMING MAX REALIZED AT Cinema ↻ LEARNING CONTINUOUS LEARNING — actual outcomes improve prediction
    Predict→ Respond→ Optimize→ Learn
    Base ForecastAudience ResponseDynamic SchedulingContinuous Learning

    Forecasting box office demand is the starting point. The cinema chain's schedules shape audience choices. Connect forecasts to scheduling decisions, then learn from actual outcomes in a closed loop.

    IE4372026.08Thank you→ Q&A · A Appendix menu · N Speaker notes
    Q&AQ&A

    We welcome your questions

    Predict→ Respond→ Optimize→ Learn
    1. The cinema chain's existing Base Demand Forecast remains a valuable asset.
    2. A 5-million-viewer forecast does not fix the cinema chain's audience share.
    3. Screen, time, frequency, and format choices change audience behavior.
    4. Prediction estimates schedule outcomes; Programming chooses the schedule.
    5. Actual booking response reoptimizes future schedules within Freeze rules.
    6. The project team integrates this intelligence into the cinema chain's architecture and evaluates it through a staged pilot.
    IE437IE437 · Classroom discussionAppendix menu — A key
    A1Appendix — “How does this improve the baseline?”

    Keep Models ① and ②;
    add Response and Optimization around them

    DomainReference processProposal — added layer
    PredictionSite/content occupancy prediction (Model ②)Maintain and improve Base Forecast + audience response to scheduling actions through Audience Response Model as a new layer
    ProgrammingApply automated scheduling rules → schedule (Model ①)Express rules as structured constraints and repeatedly call Prediction through Optimization Engine to search candidate schedules
    OperationsGenerate once → Open bookings → Review afterwardRolling Horizon reoptimization — continuously improve unopened schedules within Freeze rules
    LearningSite-level results · Schedule assessmentDecision / Outcome Log-based closed-loop retraining — actuals improve both forecast models
    SystemTheater business → EAI → DX AI → DBKeep the same paths — Intelligence Module within/alongside DX AI; preserve UI and business systems
    Build on the cinema chain's existing investmentand retain all forecasting and scheduling assets.
    APPENDIX A1
    A2Appendix — Validation gates and reusable outputs

    Validate through four gates,
    and retain useful assets at every outcome

    GATE H1

    Base Forecast

    Pass criterion At least match the current forecast on agreed metrics (MAE/WMAPE, etc.)

    Assets retained if gate fails D1 benchmark report · Data quality diagnosis

    GATE H2

    Action Response

    Pass criterion Check held-out response accuracy and supported action effects

    Assets retained if gate fails Action History roadmap · Prospective logging framework

    GATE H3

    Optimization

    Pass criterion Improve the evaluation metric over the reference, with zero hard-constraint violations

    Assets retained if gate fails D3 Engine · D4 Replay environment

    GATE H4

    Dynamic

    Pass criterion Reoptimization adds gains over the fixed initial schedule

    Assets retained if gate fails D6 operating guidelines · Draft Freeze policy

    GO / NO-GO After the integrated pilot, the decision owner decides whether to proceed to Phase 2. Deliver D6 validation report and D7 API specification regardless of the outcome.
    Every stage produces useful decision assets— retain the model, rules, comparison tools and operating lessons.
    APPENDIX A2 · Agree gate thresholds after baseline diagnosis
    A3Appendix — “Will already-booked shows change?”

    The system recommends; Cinema operators decide
    under four operating principles

    Principle 01

    Keep opened shows unchanged

    Exclude FROZEN shows from reoptimization. Handle exceptions only through existing operating procedures.

    Principle 02

    Approve near-term changes

    Propose LIMITED-window changes only for material gains. Apply only after operator approval.

    Principle 03

    Change Penalty

    Penalize schedule changes to limit disruption. Recommend changes only when gains exceed the stability threshold.

    Principle 04

    Human approval + Decision Log

    Initially, require 100% operator approval. Record every recommendation, approval, rejection, and reason.

    Customer commitments and operational stability come first— Freeze timing follows the cinema chain's operating policies.
    APPENDIX A3 · See Rolling Horizon / Freeze Window in the main deck
    A4Appendix — “What if Action History is unavailable?”

    Without Action History, adapt
    the identification strategy and validation depth

    Strategy 01

    Use natural experiments

    Use different schedules for the same film across sites/weeks as quasi-experimental variation to estimate response effects.

    Strategy 02

    Build prospective logs

    During proposed shadow operation, record schedule snapshots and revisions to strengthen identification later in the PoC and Phase 2.

    Strategy 03

    Conservative identification

    Estimate effects only where assumptions are supported; state uncertainty. Observational data alone does not establish full causality.

    Strategy 04

    Redefine validation depth

    Treat H2 as partial validation and assess PoC results around H1 and H3 (Replay-based optimization gains).

    Clear limits support credible evidence— agree the scope based on data availability at kickoff.
    APPENDIX A4 · See Action History and Required Support in the main deck
    A5Appendix — modeling candidates for technical review

    Choose methods after data diagnosis
    using clear selection criteria

    Base Forecast candidates
    • Gradient Boosting Attributes · Awareness · Advance bookings
    • Time-series models Decay by release week
    • Text embedding features Synopsis · Reviews, after utility validation
    • Ensemble benchmark Compare with Cinema forecasts on the same metrics
    Audience Response candidates
    • Hierarchical response models Site × Film × Time slot
    • Discrete-choice · Substitution models Within-film substitution · Cross-film cannibalization
    • Causal · Uplift methods Only when identification assumptions hold
    • Uncertainty estimation Prediction intervals · Calibration
    Programming candidates
    • MILP · CP exact methods Core hard constraints
    • Large Neighborhood Search Scale to large instances
    • Hybrid heuristics Evaluate nonlinear response functions
    • Rolling Horizon reoptimization Freeze · Change Penalty
    Choose the model after diagnosisusing accuracy, explainability, operating stability, and Cinema system compatibility.
    APPENDIX A5 · Method candidates for model and search selection
    A6Appendix — “Data security and export”

    Process cinema data within the cinema chain’s environment
    and exclude it from global-model training

    Principle 01

    Deidentified · Aggregated data only

    Request deidentified or aggregated film, screening and audience records. Avoid individually identifiable information.

    Principle 02

    Process in the client environment

    Plan to keep data and decision records within the cinema chain's tenant. Exclude them from provider global-model training.

    Principle 03

    Use existing paths

    Use the cinema chain's EAI · DX AI · DB paths without adding external export channels.

    Principle 04

    Deliverable ownership

    Specify operator ownership of validation reports, Decision Logs, Replay environments and project outputs. Support IT and security review.

    Complete security review before kickoff— configure VPC or on-premises processing where required by Cinema policy.
    APPENDIX A6 · Proposed data-processing principles
    A7Appendix — draft constraint catalog

    The Optimizer’s operating grammar
    Draft for agreement at kickoff

    ItemClassProposed source fieldsNotes
    Operating hours · First/last showHARDOperating hours (definition to agree)May differ by site and weekday
    Cleaning · Turnaround timeHARDCleaning time · RuntimeVaries by screen size
    Screen quotaHARDScreen quota · Domestic/foreign film categoryApplicable quota rule
    Premium-screen policy · Share capHARD SOFTPremium-screen cap · Space/technology · FormatFormat fit: Hard; share cap: Soft
    Booking lock (Freeze)HARDBooking opening time · Advance-booking scheduleRolling Horizon boundary
    Minimum / Maximum showsSOFTFilm scheduling category · Operating policyHard if required by distribution contracts
    Show spacing · Time-slot practicesSOFTOperator interviewsExtract hidden constraints
    APPENDIX A7 · Proposed constraint catalog · See Rule → Constraint in the main deck
    All worked examples Present01/36