Time-Series Forecasting for Concert Attendance: Predict Demand for Halftime Acts

Hook: Beat the deadline — forecast ticket demand and streaming spikes for headline acts

If you’ve ever stared at a messy spreadsheet of daily ticket sales or hourly stream counts and panicked before a deadline, you’re not alone. Event planners, streaming analysts and data-savvy students all face the same pain: converting noisy time-series data into reliable forecasts under tight timelines. This guide teaches two workhorse methods — ARIMA and exponential smoothing — end-to-end, using a real-world, high-stakes example: predicting ticket demand and streaming viewership around a headline halftime performance (think: Bad Bunny at the Super Bowl).

Why this matters in 2026: trends that change the forecasting game

Late 2025 and early 2026 brought three practical changes that make time-series forecasting both more powerful and more demanding:

• Greater access to near-real-time ticketing and streaming APIs from platforms and venues, enabling nowcasts and rapid re-forecasts.
• Wider adoption of probabilistic forecasting and quantile-aware models in production pipelines — decision-makers want uncertainty, not just point forecasts.
• New hybrid signals: social media momentum, influencer amplification, and trailer drops (like the kind of promo clip that hypes a halftime show) now produce measurable demand pulses that must be modeled as exogenous events.

With those trends in mind, this article gives you practical recipes and code-level ideas to forecast both ticket demand and streaming viewership before, during, and after a headline halftime act.

Quick roadmap (inverted pyramid — most important first)

1. Define the problem and the forecast horizon
2. Assemble and preprocess time-series and exogenous data
3. Fit exponential smoothing (ETS / Holt-Winters) baseline
4. Fit ARIMA / SARIMAX with exogenous regressors
5. Evaluate with rolling-origin CV and pick the model
6. Deploy probabilistic forecasts and action-ready outputs

1) Define the forecasting problem

Be explicit: what are you predicting, for whom, and on what horizon?

• Ticket demand: daily or hourly ticket purchases for the stadium (horizon: 30–90 days pre-event; hourly for last 72 hours).
• Streaming viewership: hourly concurrent streams and on-demand plays for the headliner (horizon: nowcast for the 48 hours around the halftime broadcast; 7–14 day forecasts for post-event streaming tail).

Different horizons favor different models: ETS is fast and handles seasonality well; ARIMA/SARIMAX gives more control for differencing and exogenous events. You’ll likely use both and compare.

2) Assemble and clean the data

Essential series and covariates:

• Primary series: ticket sales (time-stamped transactions aggregated by hour/day); streaming plays (hourly streams, viewers).
• Exogenous signals: promo dates (trailer drop), ad spend, social mentions, Google Trends, pre-sale windows, competitor events, weather (for in-person events).
• Calendar features: day-of-week, holiday flags, game-day schedule (kickoff time), and halftime slot timing.

Cleaning steps (practical):

1. Aggregate to the chosen frequency early (hourly vs daily).
2. Impute short gaps with interpolation; mark long gaps as missing to avoid bias.
3. Detect and annotate outliers — e.g., sudden spikes after a trailer release — rather than automatically removing them. Treat them as interventions.
4. Log-transform counts if variance grows with level; remember to bias-correct when back-transforming forecasts.

3) Exploratory analysis and stationarity

Core diagnostics you should run:

• Decompose the series (additive or multiplicative) into trend, seasonality, and residuals.
• Plot ACF and PACF to spot autoregressive or moving-average structure.
• Run unit-root tests: Augmented Dickey-Fuller (ADF) and KPSS. Use both — they have different null hypotheses.

Interpretation tips:

• If ADF fails to reject non-stationarity but KPSS rejects stationarity, you probably need differencing.
• Seasonal spikes (e.g., weekly cycles in ticket buying) suggest seasonal differencing or an ETS model with seasonal components.

4) Exponential smoothing — fast, interpretable baseline

Exponential smoothing (ETS) is a family of models that capture level, trend and seasonality via smoothing equations. It’s often the first model to try because it is fast, robust, and produces prediction intervals.

When to use ETS

• Clear seasonal patterns (daily/weekly).
• Short to medium horizons where trend and seasonality dominate.
• When you want a simple baseline or an ensemble component.

Python example (Statsmodels)

# basic ETS (additive seasonality)
from statsmodels.tsa.holtwinters import ExponentialSmoothing
model = ExponentialSmoothing(series, seasonal='add', seasonal_periods=7)
fit = model.fit()
forecast = fit.forecast(steps=14)
# get prediction intervals with simulation or statsmodels' built-in method in newer releases
  

Holt-Winters (multiplicative vs additive): choose multiplicative if seasonal amplitude grows with level (common in streaming when peaks scale with baseline).

5) ARIMA and SARIMAX — flexible, powerful, exogenous-aware

ARIMA (AutoRegressive Integrated Moving Average) models handle autoregression (AR), differencing (I) for stationarity, and moving average (MA) dynamics. For seasonality, use SARIMA or the state-space SARIMAX to include external regressors.

Key practical steps

1. Identify order (p, d, q) via ACF/PACF and unit-root tests.
2. For seasonality, identify seasonal orders (P, D, Q, m) — m is seasonal period (7 for weekly patterns in daily data, 24 for hourly data with daily seasonality).
3. Consider automatic selection with pmdarima's auto_arima to bootstrap candidate models quickly.
4. Include exogenous regressors (exog) for trailer drops, ad spend, or social volume. These often explain sudden spikes and improve forecast accuracy.

Python SARIMAX example (statsmodels)

from statsmodels.tsa.statespace.sarimax import SARIMAX
# series is a pandas Series indexed by datetime
# exog is a DataFrame with aligned datetime index containing social mentions, promo flags
model = SARIMAX(series, order=(1,1,1), seasonal_order=(1,1,1,7), exog=exog)
fit = model.fit(disp=False)
pred = fit.get_forecast(steps=14, exog=exog_future)
forecast = pred.predicted_mean
conf_int = pred.conf_int()
  

Tip: Always include the same exogenous columns in training and forecasting steps, and be careful to supply future exog values or reasonable estimates for them.

6) Handling events, interventions, and external buzz (the Bad Bunny effect)

Major promotional events — trailer drops, surprise collaborations, or an artist's viral moment — create non-repeating spikes. Treat these as:

• Interventions (indicator variables that turn on at the event and optionally decay), or
• Exogenous regressors that quantify buzz (social mentions, search volume, ad impressions).

Example intervention variable: create a promo_flag series that equals 1 on the day of trailer release and then decays exponentially for the following days to capture fading attention.

7) Model evaluation — choose the right metric and validation strategy

Standard holdout is risky for time-series. Use rolling-origin cross-validation (a.k.a. sliding window / time-series CV) to estimate realistic production performance.

Metrics

• RMSE and MAE for scale-sensitive error.
• MAPE for relative error when zeros are rare (watch out if counts include zeros).
• Pinball loss (quantile loss) for probabilistic forecasts.

Rolling-origin example (conceptual)

Split at t0, fit on [0..t0], forecast h steps, score. Then expand training window to [0..t1], repeat. Aggregate errors across folds.

8) Diagnostics and residual checks

• Plot residuals: mean near zero, constant variance, and no autocorrelation.
• Run Ljung-Box on residuals — if significant, model may miss dependencies.
• Check parameter significance and stationarity/invertibility conditions of ARIMA fits.

9) Choosing between ARIMA and ETS — practical rules

• ETS shines when seasonality and level drive the series and exogenous events are minimal.
• ARIMA/SARIMAX excels when you need to include exogenous regressors (ad spend, trailer drops) or when the residual autocorrelation implies more complex dynamics.
• Combine both via simple ensembles (mean, weighted mean) — ensembles often beat single models for noisy event-driven series.

10) Probabilistic forecasts: prediction intervals and scenarios

In 2026, decision-makers expect probabilistic forecasts. Both ETS and SARIMAX provide prediction intervals; for richer uncertainty, simulate future paths (Monte Carlo) conditioning on exogenous scenario assumptions.

Actionable scenario examples:

• Base scenario: no additional promo after trailer.
• High-buzz scenario: two major promo pushes leading to a +30% streaming uplift on game day.
• Low-buzz scenario: negative news reduces conversion rates — model with a negative exog shock.

11) Nowcasting and real-time re-forecasting

For hourly demand around halftime, set up a near-real-time pipeline that ingests ticket transactions and streaming telemetry, runs quick ETS or low-order ARIMA updates, and pushes forecasts to dashboards. Use rolling-origin re-evaluation to detect model drift.

2026 tip: automated retraining and change-point detection are more common. Use them to detect when a major promotional clip (like a halftime trailer) requires re-fitting with fresh exogenous behavior.

12) Example workflow: from raw data to action

1. Collect: hourly ticket transactions (last 120 days) + hourly streaming counts + social buzz index.
2. Aggregate and clean; create features: day-of-week, promo_flag (trailer day), cumulative_pre_sales.
3. Fit ETS for baseline daily ticket demand (7-day seasonality).
4. Fit SARIMAX for hourly streaming, with exog = social_buzz, promo_flag.
5. Run rolling-origin CV; compare RMSE and 95% coverage of intervals.
6. Produce scenario-based predictions and quantiles; feed to inventory managers and CDN capacity planners.

13) Short code recipe: full pipeline sketch (Python pseudocode)

# 1. Load data: series, exog
# 2. Train/test split with rolling-origin logic
# 3. Fit ETS baseline
ets = ExponentialSmoothing(train_series, seasonal='mul', seasonal_periods=24).fit()
ets_fore = ets.forecast(h)

# 4. Fit SARIMAX with exogenous regressors
sarimax = SARIMAX(train_series, order=(p,d,q), seasonal_order=(P,D,Q,m), exog=train_exog).fit()
sarimax_fore = sarimax.get_forecast(steps=h, exog=test_exog).predicted_mean

# 5. Evaluate
rmse_ets = rmse(ets_fore, test_series)
rmse_sarima = rmse(sarimax_fore, test_series)

# 6. Ensemble
ensemble_fore = 0.5*ets_fore + 0.5*sarimax_fore

# 7. Produce prediction intervals (simulate residuals or use model conf_int)
  

14) Real examples (interpretation)

Suppose your SARIMAX predicts a 150% hourly streaming spike during halftime with a 95% interval of 120%–180% relative to baseline. Action steps:

• For CDNs: provision upper quantile capacity (e.g., 95th) to avoid buffering if the SLA prioritizes viewer experience.
• For merch and conversion teams: prepare dynamic pricing and limited-edition bundles if ticket/stream conversions spike.
• For venue ops: anticipate rapid entry and concession demand in the halftime window if ticket scanning predicts early arrivals tied to social momentum.

15) Common pitfalls and how to avoid them

• Ignoring exogenous events: a trailer release is not noise — model it.
• Using MAPE blindly when zeros are present — prefer MAE or RMSLE where appropriate.
• Overfitting variety: complex seasonal ARIMA with many parameters can fit noise; use AIC/BIC and CV.
• Forgetting to forecast exogenous inputs: if ad spend is a predictor, you must decide what ad spend will be in the forecast window.

16) Advanced strategies and 2026-forward recommendations

• Feature-rich SARIMAX: engineer decaying promo signals, influencer coefficients (weighted by follower recency), and search interest growth rates.
• Hybrid models: combine ARIMA/ETS with ML regressors on residuals (gradient boosting) — especially effective when many exogenous features exist.
• Probabilistic ensembles: combine multiple model predictive distributions rather than point averages.
• Automated model selection with human-in-the-loop: let auto-selection suggest candidates, then validate against business logic (e.g., is a 5x spike plausible?).

Remember: a forecast is useful when it supports decisions. Supply uncertainty, test scenarios, and translate results into operational actions.

Actionable takeaways (quick checklist)

• Start with ETS as a baseline; fit ARIMA/SARIMAX when exogenous events matter.
• Always model trailer drops and promo pushes as exogenous variables or interventions.
• Use rolling-origin CV and choose metrics appropriate for counts and business needs.
• Provide prediction intervals and scenario forecasts for operations and capacity planning.
• Automate re-forecasting around key events (e.g., 48 hours before and during halftime).

Final notes: why Bad Bunny’s halftime performance is a perfect teaching case

High-profile halftime acts create concentrated, measurable demand across ticketing and streaming channels. The combination of calendar effects (the Super Bowl date), big promotional pushes (trailers and social buildup), and real-time viewer behavior (streaming spikes) makes this an ideal case to learn modern time-series workflows that matter in 2026. You practice stationarity tests, seasonal modeling, intervention analysis, exogenous feature engineering, probabilistic forecasting, and operational translation — all on a recognizably high-impact problem.

Call to action

Ready to practice? Download or spin up a notebook and try these steps on a synthetic or public data set: build an ETS baseline, fit a SARIMAX with a promo_flag, run rolling-origin CV, and produce scenario-based prediction intervals. If you want a guided walkthrough, check out our step-by-step notebooks and datasets at equations.top — start with the halftime-forecast tutorial and run the models yourself. Share your forecasts and questions so we can help you tune models for real-world decisions.

Related Reading

• How to Run Inclusive Fitness Assessments: Accessibility, Adaptive Equipment, and Data Ethics (2026)
• Green Mining Office: How Efficient Chargers and Low‑Power Macs Reduce Electricity Costs
• Building an AI-enabled Raspberry Pi 5 Quantum Testbed with the $130 AI HAT+ 2
• Psychiatry Clinic Tech Roundup 2026: From Smart Plugs to Edge AI — Devices Worth Investing In
• Viral Hiring for Quantum Engineers: Designing Puzzle-Based Recruitment Campaigns