📊 statistics

Trend and seasonality estimation for point-process time series

This paper introduces computationally simple M-estimators for trend and seasonality in point-process time series under a log-Gaussian intensity model, deriving their asymptotic properties and validating their performance through simulations and a real-world application to Chicago's bike-sharing demand.

Original authors: Daniel Gervini, Simon A. Kopischke

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniel Gervini, Simon A. Kopischke

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand the rhythm of a busy city, but instead of looking at smooth numbers on a graph, you are looking at thousands of individual, random "dots" appearing on a map or a timeline. These dots could be the exact second a bike is rented, the moment a crime happens, or a bus arriving at a stop.

This paper introduces a new way to make sense of these scattered dots when they happen over and over again, day after day.

The Problem: Too Many Dots, Too Much Noise

Usually, when statisticians look at data, they expect smooth lines. But real-world events are messy. If you look at bike rentals for just one day, you see a chaotic cloud of dots. If you look at a whole year, you see 365 of these chaotic clouds stacked on top of each other.

The authors ask: Can we find the hidden patterns in this chaos?
Specifically, they want to answer two questions:

The Trend: Is the city getting busier or quieter over the long haul? (Like a slow tide rising or falling).
The Seasonality: Is there a repeating rhythm? (Like the daily rush hour or the weekly difference between Monday and Saturday).

The Solution: The "Object-Oriented" Approach

The authors use a clever trick. Instead of trying to analyze one giant, messy year of data all at once, they treat each day as a separate "object."

Think of it like a music conductor. Instead of listening to a whole symphony at once, the conductor looks at the sheet music for Day 1, Day 2, Day 3, and so on. Even though the musicians (the dots) are playing randomly, the conductor knows that Day 1, Day 2, and Day 3 are all part of the same song.

By treating each day as a distinct object in a time series, they can use simpler math to find the underlying melody (the trend) and the repeating chorus (the seasonality).

The "Ghost" Intensity

The paper assumes that behind every day's random dots, there is an invisible "ghost" map that dictates where and when the dots should appear. This is called the intensity function.

The Analogy: Imagine a foggy night where you can't see the streetlights clearly, but you can see the moths flying around them. The moths are the random dots (bike rentals). The streetlights are the invisible "ghost" intensity.
The authors' method tries to reconstruct the shape and brightness of those streetlights, even though the moths are buzzing around randomly.

They break this "ghost" light down into three parts:

The Baseline: The general shape of the streetlight (e.g., it's always brighter in the middle of the day).
The Trend: The lightbulb getting brighter or dimmer over the year (e.g., more bikes in summer, fewer in winter).
The Seasonality: The light flickering in a specific pattern (e.g., brighter on weekdays, dimmer on weekends).

The Chicago Bike Test

To prove their method works, the authors looked at real data from Chicago's Divvy bike-sharing system. They picked three different neighborhoods:

A Residential Area: Like a quiet suburb.
The Downtown Loop: The busy city center.
A Tourist Spot: Near the aquarium.

What they found:

The Trend: All three spots got busier in the summer and quieter in the winter, just like the temperature. But the amount of change was different. The tourist spot exploded in popularity in summer (30 times more bikes!), while the downtown spot only grew a little (3 times).
The Seasonality (The Rhythm):
- Residential: Two big spikes on weekdays (morning and evening commute) and one small spike on weekends (lunchtime).
- Downtown: One huge spike on weekdays (people going home at 6 PM) and almost nothing on weekends.
- Tourist: One big spike every day at 3 PM, but even bigger on weekends.

Why This Matters

The paper claims that previous methods couldn't handle this kind of data well because the dots are too random and the days aren't independent (today's traffic affects tomorrow's).

Their new method is like a smart filter. It ignores the random noise of individual bike rentals and reveals the clear, predictable patterns underneath. It allows city planners to see exactly when and where bikes will be needed, not just on average, but with a clear picture of the daily and yearly rhythms.

In short, the authors built a mathematical tool that turns a chaotic cloud of random dots into a clear, readable story about how a city moves.

Technical Summary: Trend and Seasonality Estimation for Point-Process Time Series

Problem Statement
The paper addresses the statistical challenge of estimating trend and seasonality in time series of point processes. While point-process modeling is well-established for single realizations (spatial, temporal, or spatio-temporal), the literature lacks robust methods for analyzing replicated point processes that form a time series. In many applications, such as slicing a long annual process into daily intervals, the resulting replications are not independent but exhibit temporal dependence. Existing functional data analysis (FDA) methods are unsuitable because they assume smooth functions, whereas point-process replications are inherently discrete and random. Furthermore, standard time-series differencing techniques cannot be applied to point processes, as the difference between two sets of points is not statistically well-defined. The authors aim to fill this gap by proposing estimators for trend and seasonal components in a time series of point processes.

Methodology
The authors propose a framework based on object-oriented data analysis, treating a collection of point processes $\{X_t : t = 1, \dots, n\}$ as the primary objects of study.

Modeling Framework:
- The data are modeled as a sequence of doubly-stochastic Poisson (Cox) processes.
- The intensity function $\Lambda_t(u)$ is assumed to be log-Gaussian: $\log \Lambda_t(u) = \mu(u, t) + Y_t(u)$ , where $Y_t(u)$ is a zero-mean Gaussian process representing latent variability.
- The mean function $\mu(u, t)$ $μ (u, t)$ is decomposed additively to capture trend and seasonality.
  - Trend-only model: $\mu(u, t) = \mu(u) + c(t)$ .
  - Trend-plus-seasonality model: $\mu(u, t) = \mu(u) + c(t) + s_{j(t)}(u)$ , where $j(t)$ indicates the seasonal period (e.g., day of the week).
- The trend $c(t)$ and seasonal components $s_j(u)$ are parameterized using splines (piecewise polynomials). The trend is modeled as a global polynomial (or spline) over a periodic interval, while spatial/temporal basis functions (e.g., B-splines) model the intensity shapes.
Estimation Approach:
- The authors employ M-estimation (specifically, a working-likelihood approach) rather than full maximum likelihood.
- The estimators are derived by maximizing a working likelihood function $\rho_n$ that assumes the latent Gaussian process $Y_t(u)$ is zero (i.e., assuming a standard Poisson process), even though the true model includes latent variability.
- This approach avoids the need to specify or estimate complex autocovariance structures for the latent processes, making the estimators computationally simple. The objective functions are shown to be concave, ensuring straightforward optimization.
Asymptotic Theory:
- The paper derives the asymptotic consistency and normality of the estimators as the number of replications $n \to \infty$ .
- Under specific assumptions (including stationarity of the latent process and decay of autocovariances), the estimators converge to the true parameters (adjusted for the variance of the latent process).
- The asymptotic distribution follows a "sandwich" variance formula ( $\Omega = W^{-1}VW^{-1}$ ), typical for M-estimators, allowing for the construction of confidence bands for the trend and seasonal intensity functions.

Key Results

Simulation Study: The finite-sample performance was evaluated via simulations with sample sizes $n=50, 100, 300$ $n = 50, 100, 300$ . The study considered three scenarios for the latent process: no variability ( $Y_t=0$ $Y_{t} = 0$ ), independent variability, and autoregressive variability.
- Results indicate that the estimators are consistent: bias and root mean squared error (RMSE) decrease as $n$ increases.
- Standard deviations dominate the error, and autocorrelation in the latent process (autoregressive case) increases the standard deviation due to reduced effective sample size, but the estimators remain consistent.
Real-Data Application: The method was applied to the Divvy bike-sharing system in Chicago (2016 data).
- The study analyzed check-out times at three distinct stations (residential, downtown, and touristic) as temporal point processes.
- The model successfully captured a weekly seasonality ( $d=7$ ) and an annual trend driven by temperature.
- Findings:
  - Trend Magnitude: The annual trend varied significantly by location. The touristic station (Shedd Aquarium) showed a 30-fold increase in summer demand compared to winter, whereas the downtown station showed only a 3-fold increase.
  - Seasonal Patterns: Distinct demand profiles were identified. The residential station showed morning and evening commute peaks on weekdays and a single noon peak on weekends. The downtown station showed a single evening peak on weekdays and low, flat demand on weekends. The touristic station showed a single afternoon peak (3 PM) for all days, higher on weekends.

Significance and Claims
The paper claims to introduce a new class of log-Gaussian models tailored for replicated point processes, distinct from existing single-realization Cox process literature. Its primary significance lies in:

Methodological Novelty: Providing a simpler, object-oriented alternative to complex point-process techniques, enabling the use of standard statistical inference concepts (like trend and seasonality decomposition) for point-process time series.
Computational Simplicity: The use of M-estimators with working likelihoods avoids the computational burden of modeling latent autocovariance structures directly.
Practical Utility: The ability to decompose complex spatiotemporal patterns (like bike demand) into interpretable trend and seasonal components without requiring high-frequency smoothing or large counts per replication.

The authors explicitly note that while the methods are based on splines and can be applied to both temporal and spatial settings, the current work focuses on the theoretical derivation and application to temporal examples. They do not claim to eliminate the need for modeling latent dependencies entirely but rather provide a robust estimation framework that remains valid even when the working model (independence) is misspecified.