A Bayesian-Inference Method for Measuring Box Office Outperformance

Chinese

The film industry is a market with exceptionally high uncertainty, and box office performance directly affects both the financial condition and market valuation of production companies. The degree of box office outperformance is a key indicator of how well a movie is received by the market and an important variable in evaluating film-company performance. Traditional box office forecasting methods, however, often rely on static point-estimate models. They struggle to capture the dynamic evolution of box office revenue over time, and they are even less capable of quantifying how far box office results exceed market expectations and with what level of uncertainty. As a result, they are difficult to convert into actionable investment strategies.

Why This Matters

This article proposes a Bayesian-inference method for measuring box office outperformance. By constructing a Bayesian time series model based on a trend-decay dynamic mechanism, the method combines human priors with real-time box office data to produce probabilistic box office forecasts and quantify the degree of outperformance. The method can both improve forecast accuracy and quantify the probability of outperforming expectations, providing a more interpretable measurement tool.

The core advantages of Bayesian methods in box office forecasting are:

1. They provide a full posterior distribution, allowing us to calculate the probability of box office outperformance rather than only a single-point forecast.
2. They update forecasts dynamically, adapting to market changes, whereas traditional models must be retrained.
3. They quantify uncertainty explicitly, providing confidence intervals and probabilistic interpretations instead of one single predicted value.

A Bayesian Framework for Box Office Forecasting

Let \(y_t\) denote the box office revenue of a film on day \(t\), and let \(\mathbf{x}_t\) denote the features affecting box office performance, such as marketing spend or audience ratings. The objective is to forecast \(y_t\) and evaluate the degree to which it exceeds expectations.

(1) Prior distribution:
The Bayesian model first assumes that the parameter vector \(\theta\) follows a prior distribution, which can be set by the researcher based on market forecasts or subjective judgment:

$$ p(\theta) \sim \mathcal{N}(\mu_0, \sigma_0^2) $$

Here, \(\theta\) represents key box office parameters such as the growth rate and decay coefficient.

(2) Likelihood function:
The observed data \(y_t\) follow a noisy distribution:

$$ y_t \mid \theta, \mathbf{x}_t \sim \mathcal{N}(f(\mathbf{x}_t, \theta), \sigma^2) $$

Here, \(f(\mathbf{x}_t, \theta)\) is the mean forecasting function for box office performance. It may be linear or nonlinear, such as a neural network or a Gamma regression model.

(3) Posterior update:
Using Bayes’ theorem, the posterior distribution of \(\theta\) is updated from observed data:

$$ p(\theta \mid y_{1:t}) \propto p(y_{1:t} \mid \theta) p(\theta) $$

The future box office value \(y_{t+1}\) is then forecast as

$$ p(y_{t+1} \mid y_{1:t}) = \int p(y_{t+1} \mid \theta) p(\theta \mid y_{1:t}) d\theta $$

At this point, what we obtain is a distribution, not a single forecast value.

The Mathematical Superiority of the Bayesian Approach

Predictive Uncertainty

Traditional methods such as regression provide only a point estimate:

$$ \hat{y}_t = f(\mathbf{x}_t; \hat{\theta}) $$

whereas a Bayesian approach provides the full posterior predictive distribution:

$$ p(y_t \mid y_{1:t-1}) $$

This allows us to estimate **confidence intervals**, for example the 95% predictive interval:

$$ [\mathbb{E}[y_t] - 1.96 \sigma, \mathbb{E}[y_t] + 1.96 \sigma] $$

This is particularly important for highly volatile box office data.

Dynamic Updating to Adapt to Market Change

Bayesian inference allows us to update forecasts dynamically as new data arrive:

$$ p(\theta \mid y_{1:t}) \rightarrow p(\theta \mid y_{1:t+1}) $$

As a result, the model can adapt to new trends such as:

• A film’s word of mouth gradually gaining traction and producing box office outperformance
• A sudden social-media event boosting audience interest

Traditional methods, by contrast, usually require retraining and therefore cannot adapt in real time.

Interpretability

Bayesian methods allow us to calculate directly:

$$ P(y_t > \mathbb{E}[y_t]) $$

Compared with black-box machine learning methods such as deep learning, this gives a more intuitive probabilistic interpretation. For example:

• “The probability that this movie’s box office today exceeds the forecast is 98%” -> indicating that market reception is far stronger than expected
• “The probability that the box office falls below the 90% confidence interval is only 5%” -> indicating that the model forecast is robust

Box Office Time-Series Modeling

Movie box office revenue usually exhibits a typical life-cycle pattern, which can be decomposed into a growth phase and a decay phase. We construct the following dynamic model:

$$ y(t;S_0,\mu, \lambda) = S_0 \cdot (\mu t + 1) \cdot e^{-\lambda t} + \epsilon_t $$

where:

• \(S_0\): the opening-day box office baseline, determined by factors such as production scale and IP influence
• \(\mu\): the growth-rate parameter, reflecting the effect of word-of-mouth diffusion
• \(\lambda\): the decay-rate parameter, reflecting declining audience interest
• \(\epsilon_t \sim \mathcal{N}(0,\sigma^2)\): observation noise

This model has a clear economic interpretation:

1. The term \((\mu t +1)\) captures word-of-mouth-driven linear growth, such as diffusion through social media
2. The term \(e^{-\lambda t}\) reflects natural decay, with audience interest diminishing over time
3. The multiplicative form guarantees a single-peaked shape consistent with real box office curves

Fitting the Time-Series Model

Using box office event series from more than 280 films on the Maoyan Movie platform, we fit the parameters film by film to test whether the time-series model captures the general trend in box office dynamics.

For a given film, the fitting problem reduces to solving the optimization problem

$$ \theta = \mathrm{argmin}_{\theta}\ \mathcal{L}(\theta),\quad \mathcal{L}(\theta) = \frac{1}{T}\sum_t (\hat{y}_t(\theta) - y_t)^2 $$

where \(T\) is the length of the observed box office time series, \(y_t\) is the day-\(t\) box office value from Maoyan, and \(\hat{y}_t\) is the model’s day-\(t\) box office forecast under parameter \(\theta\).

By solving this optimization problem for each film, we obtain the corresponding best-fit parameter set \(Y ={(S_0^{(i)},\mu^{(i)},\lambda^{(i)})}_{i=1}^n\).

We then calculate the mean and variance of the linear growth parameter and natural decay parameter:

$$ m_{\mu} = \frac{1}{n}\sum_i \mu^{(i)},\quad\sigma_{\mu}^2 = \frac{1}{n}\sum_i(\mu^{(i)} - m_{\mu})^2 $$

$$ m_{\lambda} = \frac{1}{n}\sum_i \lambda^{(i)},\quad\sigma_{\lambda}^2 = \frac{1}{n}\sum_i(\lambda^{(i)} - m_{\lambda})^2 $$

A LightGBM Prior-Generation Model

Model Setup

Using an ensemble-learning approach, we train a model on pre-release film features to predict the opening-day box office parameter and then use that prediction to construct the prior distribution of the parameters.

Training data and target variable:

• Feature matrix \(X \in \mathbb{R}^{n \times p}\), where there are \(n\) historical films and \(p\) features
• Target variable \(Y ={S_0^{(i)}}_{i=1}^n\), the opening-day box office parameters for the \(n\) films

Using LightGBM to predict the opening-day box office parameter

For a new film with features \(x_{\text{new}}\), the ensemble model yields the following opening-day box office forecast:

$$ \hat{S}_0 = \mathrm{model}(x_{\mathrm{new}}) $$

Bayesian Parameter Estimation

Assign conjugate priors to the parameters \(\theta = (S_0,\mu,\lambda)\):

$$ \begin{aligned} p(S_0) =& \mathcal{N}^+(S_0;\hat{S}_0,\sigma_S^2) \\ p(\mu) =& \mathcal{N}^+(\mu; m_{\mu},\sigma^2_{\mu}) \\ p(\lambda)=& \mathcal{N}^+(\lambda;m_\lambda,\sigma_\lambda^2) \end{aligned} $$

Here \(\mathcal{N}^+\) denotes a truncated normal distribution, ensuring that all three parameters are nonnegative. On top of that, users may incorporate additional prior information from media forecasts:

$$ \begin{aligned} p'(S_0) =& \mathcal{N}^+(S_0;\hat{S}_0',\sigma_S'^2) \\ p'(\mu) =& \mathcal{N}^+(\mu; m_{\mu}',\sigma'^2_{\mu}) \\ p'(\lambda)=& \mathcal{N}^+(\lambda;m_\lambda',\sigma_\lambda'^2) \end{aligned} $$

Combining the model-generated prior with the user-defined prior gives the final prior distribution:

$$ \mathrm{prior}(\theta) =p(S_0)p(\mu)p(\lambda)\times p'(S_0)p'(\mu)p'(\lambda) $$

Given the observed post-release box office sequence \({y_t}_{t=1}^T\), we construct the likelihood function based on the maximum entropy principle:

$$ \mathrm{likelyhood}(\theta;y_{1:T}) = \Pi_i \frac{1}{\sqrt{2\pi \sigma_i^2}} \exp{\left(-\frac{(y_t-\hat{y}_t(\theta))^2}{2\sigma_i^2}\right)} $$

Combining the parameter priors with the likelihood gives the posterior distribution of the parameters. Markov chain Monte Carlo (MCMC) sampling is then used to approximate the posterior \(p(\theta|y_{1:T})\):

$$ p(\theta|y_{1:T}) = \mathrm{prior}(\theta)\times \mathrm{likelyhood}(\theta;y_{1:T}) $$

Based on these samples, we can forecast the box office time series, total box office, and other important quantities while providing their probability intervals.

Measuring the Degree of “Outperformance”

Total Box Office

The posterior distribution of total box office \(Y = \sum_{t=1}^T y_t\) can be approximated by Monte Carlo integration:

1. Draw \(N\) parameter samples \(\theta^{(i)}\) from the posterior \(p(\theta|y_{1:t})\)

2. For each parameter draw, compute

$$ Y^{(i)} = \frac{S_0}{1 - e^{-\lambda}} \left( 1 + \frac{\mu e^{-\lambda}}{1 - e^{-\lambda}} \right) $$

3. Obtain the empirical distribution \({Y^{(i)}}_{i=1}^N\) for total box office

Define “outperformance” in two quantitative ways:

1. Absolute-threshold method
Given an industry forecast \(Y_{\text{ref}}\), such as the Maoyan Pro estimate, compute the probability of beating it:

$$ P_{\text{beat}} = \frac{1}{N}\sum_{i=1}^N \mathbb{I}(Y^{(i)} > Y_{\text{ref}}) $$

2. Relative-quantile method
Using the median of the model’s own forecast distribution, \(Y_{\text{med}}\), as the benchmark, compute:

$$ P_{\text{extreme}} = P(Y > Y_{\text{med}} + k\sigma_Y) $$

Here \(\sigma_Y\) is the posterior standard deviation, and \(k\) can be adjusted according to risk preference, usually between 1 and 2.

Quantifying Word of Mouth Through Parameter Deviations

Movie word of mouth can be quantified scientifically by measuring the deviation between posterior parameters and prior expectations, leading to the following evaluation system.

Word-of-Mouth Strength Indicators

(1) Growth-Parameter Deviation (\(\mu\)-Deviation)

Definition:

$$ D_\mu = \frac{\mu_{\text{post}} - \mu_{\text{prior}}}{\sigma_{\mu_{\text{prior}}}} $$

• \(D_\mu > 1.64\): significantly positive word of mouth (\(p<0.05\))
• \(D_\mu < -1.64\): significantly negative word of mouth

(2) Decay-Parameter Deviation (\(\lambda\)-Deviation)

$$ D_\lambda = \frac{\lambda_{\text{prior}} - \lambda_{\text{post}}}{\sigma_{\lambda_{\text{prior}}}} $$

• \(D_\lambda > 0\): lifecycle extension
• \(D_\lambda < 0\): rapid deterioration

Composite Word-of-Mouth Index

Construct a weighted indicator:

$$ WOM = w_\mu D_\mu + w_\lambda D_\lambda $$

Suggested weights:

• \(w_\mu = 0.6\) (the growth effect matters more)
• \(w_\lambda = 0.4\)

Rating criteria:

• WOM > 2: breakout-level word of mouth
• 1 < WOM ≤ 2: strong word of mouth
• |WOM| ≤ 1: in line with expectations
• WOM < -1: word-of-mouth collapse