Advanced Analytics · Forecasting · Multiplier Analysis · BigQuery · Python

Admission Multiplier Analysis & Forecasting Experiment

As part of my Data Analyst Internship, I explored analytical methods to estimate actual cinema admission from tracked scraping data and experimented with forecasting approaches for predicting daily filled seats. This project combined multiplier binning analysis, AVG OR XXI patterns, LightGBM experimentation, Synthetic Control Method, donor weighting, and hybrid sequential prediction.

Project Overview

In cinema admission monitoring, the number of Filled Seats captured from scraping data does not always perfectly represent the actual admission number. Some ticket purchases may happen through offline counters or channels that are not fully reflected in the scraped data. To estimate actual admission more realistically, a multiplier is often used to adjust tracked admission values.

This project explored two related analytical directions. First, I analyzed the relationship between AVG OR XXI and historical admission multipliers using a binning method. Second, I experimented with forecasting daily filled seats using Synthetic Control Method and several donor weighting techniques such as Ridge, Lasso, ElasticNet, and XGBoost.

The project was treated as an analytical experiment rather than a final production model. The goal was to understand the patterns, limitations, and potential approaches for estimating actual admission and predicting future cinema performance.

Business Problem

The main challenge was that tracked admission data from scraping may underrepresent actual ticket sales. This creates uncertainty when business users need to interpret daily movie performance or estimate actual admission numbers.

• Tracked admission may be lower than actual admission: Not all purchases are fully captured through scraping.
• Multiplier values are not constant: The multiplier can change depending on movie demand, day type, audience behavior, and scraping coverage.
• Weekday and weekend behavior can differ: On weekends or holidays, more people may buy tickets earlier online, making tracked data closer to actual performance.
• Direct forecasting is difficult: Daily filled seats can be highly volatile and influenced by show allocation, competition, release phase, and audience momentum.
• Comparable historical movies may not always exist: Forecasting methods such as Synthetic Control Method depend heavily on finding suitable donor movies.

Analytical Approach

The project used two main approaches: multiplier binning and forecasting experimentation. The multiplier analysis was designed to provide an interpretable reference range for daily adjustment, while the forecasting experiment tested whether historical movie patterns could help predict future filled seats.

Admission Multiplier Analysis

A multiplier is a factor used to estimate actual admission from tracked filled seats. If the tracked admission is lower than actual admission, the multiplier helps adjust the tracked number closer to the expected actual value.

Estimated Actual Admission =
  Total Tracked Filled Seats × Multiplier

The analysis used AVG OR XXI as the main indicator because XXI has wide cinema coverage and a large number of showtimes. This makes its occupancy rate useful as a proxy for national-level demand movement.

Multiplier Assumption

The initial assumption was that audience ticket-purchasing behavior changes depending on movie demand and day type. On weekends, holidays, or high-demand days, more people tend to buy tickets earlier online because they are worried about missing seats. As a result, tracked data tends to be closer to actual admission, and the multiplier may be lower.

On regular weekdays, more people may buy tickets closer to showtime or directly at the counter. This can cause tracked data to be lower than actual admission, meaning the multiplier may need to be higher.

Binning Method

The main method used for multiplier analysis was binning. Binning groups AVG OR XXI into several ranges, then calculates the average, minimum, and maximum multiplier within each range. This makes the output easier to interpret and more practical for daily monitoring.

Instead of relying on a black-box model, the binning method gives users a readable reference:

Multiplier Pattern by AVG OR XXI Bin =
  Group AVG OR XXI into ranges
  → Calculate average multiplier
  → Calculate minimum multiplier
  → Calculate maximum multiplier
  → Use as daily multiplier reference

LightGBM Multiplier Prediction Experiment

In addition to binning, I also explored a LightGBM-based prediction approach to estimate multiplier values automatically. The model used several time-based and lag features, such as whether the date was a weekend, previous-day AVG OR XXI, previous-day multiplier, and previous-day tracked admission.

However, for daily usage, the binning method was considered more practical because it is easier to explain, easier to validate, and more flexible when combined with business intuition.

Forecasting Experiment with Synthetic Control Method

The second part of the project explored whether historical movie patterns could be used to predict future daily filled seats. The main concept tested was Synthetic Control Method, where a target movie is compared to a weighted combination of historical donor movies.

In theory, if the donor movies have similar patterns to the target movie, their weighted combination can form a synthetic trajectory that approximates the target movie's future performance.

Donor Selection & Covariates

Donor movies were selected based on similarity to the target movie. The similarity matrix used several covariates that describe release scale, audience behavior, showtime distribution, competition, and market composition.

Donor Weighting Methods

Several methods were tested to calculate donor weights. Each method used the same general pipeline but differed in how it assigned importance to donor movies.

Hybrid Sequential Prediction

The final prediction was not based only on the synthetic trajectory. A hybrid sequential prediction approach was tested to make the forecast more adaptive. This approach combined donor-based movement with business rules and temporal patterns.

• Daily chain projection: follows daily movement from the synthetic donor trajectory.
• Week-over-week baseline: compares the target movie against the same day from the previous week.
• Day-of-week correction: adjusts for weekday and weekend admission patterns.
• Phase detection: identifies whether the movie is in a growth, decline, or stable phase.
• Adaptive blending: combines daily projection and week-over-week projection based on current phase.
• Safety clipping: prevents predictions from becoming unrealistically high or low.

Hybrid Prediction =
  Blend(
    Daily Chain Projection,
    Week-over-Week Baseline,
    Day-of-Week Correction,
    Phase Detection,
    Safety Clipping
  )

Selected Analytical Logic

The following simplified and masked pseudo-code shows the overall logic used for multiplier binning and forecasting experimentation.

# Simplified logic for portfolio publication
# Sensitive project identifiers are masked.

# 1. Load historical screening data from BigQuery
screening_data = load_bigquery_table(
    project="masked-gcp-project",
    dataset="masked_dataset",
    table="Fact_Screening"
)

# 2. Create daily admission aggregation
daily_sales = (
    screening_data
    .groupby(["movie_title", "show_date"])
    .agg({
        "Filled_Seats": "sum",
        "Total_Seats": "sum"
    })
    .reset_index()
)

# 3. Multiplier binning analysis
or_xxi_daily = calculate_avg_or_xxi(daily_sales)
multiplier_history = load_multiplier_reference()

multiplier_dataset = merge_by_date(
    or_xxi_daily,
    multiplier_history
)

multiplier_dataset["or_bin"] = create_quantile_bins(
    multiplier_dataset["avg_or_xxi"]
)

multiplier_reference = (
    multiplier_dataset
    .groupby("or_bin")
    .agg({
        "multiplier": ["count", "mean", "min", "max"]
    })
)

# 4. Forecasting experiment with donor movies
target_movie = "TARGET_MOVIE"

covariates = build_movie_covariates(daily_sales)
donors = select_similar_donors(
    target_movie=target_movie,
    covariates=covariates,
    method="nearest_distance"
)

# 5. Donor weighting experiment
weights = fit_donor_weighting_model(
    target_movie=target_movie,
    donors=donors,
    method="Ridge / Lasso / ElasticNet / XGBoost"
)

synthetic_trajectory = combine_donor_trajectories(
    donors=donors,
    weights=weights
)

# 6. Hybrid sequential prediction
prediction = hybrid_prediction(
    target_actual=daily_sales,
    synthetic_trajectory=synthetic_trajectory,
    use_daily_chain=True,
    use_week_over_week=True,
    use_day_of_week_correction=True,
    use_phase_detection=True,
    use_safety_clipping=True
)

evaluate_actual_vs_prediction(prediction)

Results & Limitations

The multiplier binning method was the most practical output from this project because it produced an interpretable range of multiplier values based on AVG OR XXI. It was easier to explain to business users and could be combined with domain knowledge or daily monitoring intuition.

The forecasting experiments showed that predicting daily filled seats is much more challenging. The Synthetic Control Method and hybrid donor weighting approach produced results that were still unstable for some target movies. The main limitation was donor quality: if the available historical donor movies were not sufficiently similar to the target movie, the synthetic trajectory became less representative.

Output & Impact

• Created a multiplier reference framework based on AVG OR XXI binning.
• Helped estimate actual admission from tracked scraping data.
• Provided average, minimum, and maximum multiplier references for different OR ranges.
• Explored LightGBM as an alternative multiplier prediction approach.
• Built an experimental forecasting pipeline using Synthetic Control Method.
• Compared donor weighting methods such as Ridge, Lasso, ElasticNet, and XGBoost.
• Tested hybrid sequential prediction using daily chain, week-over-week baseline, day-of-week correction, and safety clipping.
• Identified limitations of SCM when similar historical donor movies are not available.

Tools & Technologies

Key Learnings

This project helped me understand that not every analytical problem needs to be solved with a complex predictive model. In some operational contexts, a transparent binning-based approach can be more practical and easier for business users to interpret.

I also learned that forecasting cinema admission is highly sensitive to movie-specific behavior, show allocation, competition, release phase, and audience momentum. Even advanced methods such as Synthetic Control Method can become unstable when the available donor movies are not sufficiently similar to the target movie.

Another key learning was the importance of communicating model limitations honestly. The forecasting experiment provided useful technical insights, but the results still required validation and should not be treated as a final automated decision system without further development.