Forecasting Methodology

Mathematical foundations of PaceCast: why we reject MAPE, benchmark every run against a seasonal repetition baseline, and calibrate distribution-free conformal intervals.

Last updated: October 2026

Many AI analytics tools claim high accuracy using opaque, proprietary metrics or flawed formulas like MAPE that produce mathematical singularities on days with low conversion counts. PaceCast adheres to strict, empirical statistical principles. This document details the mathematical foundations behind our models.

Error Metric: Why WAPE Over MAPE

Standard Mean Absolute Percentage Error (MAPE) calculates error on each day as a ratio of that day's actual value:

MAPE = (1 / n) · ∑ | actual_t − predicted_t | / actual_t

In advertising datasets, small campaigns often have days with 0 or 1 conversion. If actual is 1 and predicted is 3, MAPE reports a 200% error. If actual is 0, MAPE divides by zero and produces an undefined infinity. A single quiet day can distort the entire monthly score.

PaceCast instead uses Weighted Absolute Percentage Error (WAPE), also known as MAD/Mean ratio:

WAPE = ( ∑ | actual_t − predicted_t | ) / ( ∑ actual_t )

WAPE divides total absolute unit errors by total actual volume across the evaluation window. High-volume shopping days contribute proportionally to the error score, while low-volume days do not destabilize the calculation.

The 7-Day Seasonal-Naive Benchmark

An accuracy score of 12% WAPE is meaningless in a vacuum. If a simple calendar rule achieves 11%, a complex machine learning model is adding noise rather than signal.

On every analysis run, PaceCast computes a parallel Seasonal-Naive benchmark:

predicted_naive(t) = actual(t − 7)

This benchmark assumes that next Monday will exactly equal last Monday. In our report, we present both WAPE scores side-by-side, along with the percentage predictive lift:

Predictive Lift = ( WAPE_naive − WAPE_model ) / WAPE_naive

Split Conformal Prediction Intervals

Classical statistical confidence intervals assume Gaussian (normal) errors and homoscedastic (constant) variance. Real ad auction numbers are skewed, zero-inflated, and subject to volatility spikes.

PaceCast utilizes inductive split conformal prediction to establish distribution-free 90% prediction intervals:

  1. Calculate absolute residual nonconformity scores on validation folds: s_i = | y_i − ŷ_i |
  2. Compute the empirical (1 − α) quantile where α = 0.10:
    q = Quantile( s, ceil( (n + 1)(1 − α) ) / n )
  3. Form the future prediction interval for any horizon day:
    C(x) = [ max(0, ŷ − q), ŷ + q ]

5-Fold Rolling-Origin Backtesting

Shuffled k-fold cross validation is invalid for time-series forecasting because training on future days leaks trend information into the past.

PaceCast executes rolling-origin backtesting across 5 chronological splits. Each split simulates an exact historical moment in time, trims the 7-day attribution embargo window, forecasts 30 days ahead, and scores errors strictly against withheld actuals.

The 7-Day Attribution Embargo

Because ad networks credit conversions back to the original click date, the most recent days in any fresh dataset suffer from conversion lag. Training on these immature cohorts causes machine learning models to project a false efficiency collapse.

PaceCast automatically excludes the trailing 7 days from model training, ensuring only mature cohorts inform baseline parameters.