Forecasting methods.
Forecasting registration, attendance, revenue and space. Curve fitting, sparse series, cold start for a new event, hierarchical forecasts, prediction intervals and backtesting.
20 articles · Forecasting methods
- How the additive pickup model forecasts your final registration numberadditive pickup modelThe additive pickup model adds an average historical increment to the registrations in hand. The arithmetic, the spread around it, and the two ways it breaks.
- The multiplicative pickup model scales your count by a historical ratiomultiplicative pickup modelThe multiplicative pickup model divides the registrations in hand by a historical share of final. Worked on a growing show, with the horizon where it inflates.
- Pickup model diagnostics that tell you which form to trust this yearpickup model diagnosticsPickup model diagnostics settle the additive versus multiplicative argument with residuals from your own closed editions instead of an opinion in the meeting.
- Fitting a booking curve to five editions with a three parameter logisticfitting a booking curveFitting a booking curve with a three parameter logistic: borrow the midpoint and steepness from closed editions, then solve the asymptote from the live one.
- Monotone regression splines keep a cumulative registration curve from going backwardsmonotone regression splinesCumulative registrations never fall, so a smoother that dips is wrong. Monotone regression splines and the PAVA pass impose the constraint before you forecast.
- Penalised spline forecasting for pacing curves that only have five editions behind thempenalised spline forecastingPenalised spline forecasting puts many knots along the days-to-open axis and lets one penalty do the work. How to pick that penalty when you have five editions.
- Booking curve principal components compress a decade of editions into two numbersbooking curve principal componentsBooking curve principal components turn each edition into two scores. What the components mean, how to fit them, and where the level has to come from.
- The Gompertz growth curve fits a registration campaign the logistic cannotgompertz growth curveThe Gompertz growth curve puts its inflection at 37 per cent of the asymptote instead of 50, which is why it fits an asymmetric registration campaign.
- Days to open as the axis when a show moves in the calendardays to open axisA days to open axis re-indexes every edition so day zero is show open, which lets a show that moved from March to June be compared with its own history.
- Aligning booking windows when one edition opened registration eleven weeks latealigning booking windowsAligning booking windows when editions had different runways: rescale to fraction of window elapsed, or truncate, and the halfway test that decides which.
- Forecasting with five observations and the parameter budget that leaves youforecasting with five observationsForecasting with five observations is a budgeting problem. Count what each method wants to estimate, then pick the one your five editions can actually pay for.
- Overfitting short time series is easy when every model fits five pointsoverfitting short time seriesOverfitting short time series is the default with five editions. A quadratic explains 98 per cent of the variation and invents 4,600 registrations of growth.
- A shrinkage estimator pulls one noisy show forecast toward the portfolio meanshrinkage estimatorA shrinkage estimator moves one show's 22 per cent growth toward the portfolio's 6 per cent. The weight is between-show variance over total variance, here 0.40.
- Borrowing strength across shows when no single show has enough historyborrowing strength across showsBorrowing strength across shows gives each show its own pooling weight. A three edition show keeps 43 per cent of its own curve, a twelve edition show 75.
- A global forecasting model trained on every show in the portfolio at onceglobal forecasting modelA global forecasting model fits one function across every show. The budget that buys three weeks of memory locally buys twenty four globally.
- Pooled panel regression turns eight shows and five years into forty rowspooled panel regressionPooled panel regression stacks eight shows and five years into forty rows. Show and year effects cost twelve of them before a single covariate is estimated.
- Hierarchical forecast reconciliation makes show and portfolio numbers add uphierarchical forecast reconciliationHierarchical forecast reconciliation projects base forecasts onto the coherent space, so a 500 registration gap between show sum and portfolio total disappears.
- Bottom up versus top down forecasting for a portfolio of eight showsbottom up versus top downBottom up versus top down decides which level you trust. On a show whose share drifted from 3.1 to 7.1 per cent, static proportions come out 36 per cent low.
- Temporal hierarchy forecasting reconciles the weekly pace with the annual numbertemporal hierarchy forecastingTemporal hierarchy forecasting aggregates one series to several frequencies, forecasts each, then reconciles. A 710 registration disagreement settles at 14,370.
- Prediction interval coverage is worse than the label claims with five editionsprediction interval coveragePrediction interval coverage on five editions is nothing like 95 per cent. The correct interval on this worked example is 2.35 times as wide as published.