Forecasting as Rendering: A 2D Gaussian Splatting Framework for Time Series Forecasting

This paper introduces TimeGS, a novel time series forecasting framework that reframes prediction as 2D generative rendering by leveraging adaptive Gaussian kernels and a chronologically continuous rasterization mechanism to overcome the topological mismatches and resolution inefficiencies of existing 2D reshaping methods, thereby achieving state-of-the-art performance.

The Gist

Imagine you are trying to predict the weather for the next week. You have a long list of past temperatures, one number for every hour. This is called Time Series Forecasting.

Most computer programs try to solve this by looking at the list as a long, straight line of numbers. They try to guess the next number based on the previous ones. But real-world data (like weather, electricity usage, or traffic) isn't just a straight line; it has cycles (day/night, weekly patterns) and sudden spikes (storms, rush hour).

This paper introduces a new way to think about this problem, called TimeGS. Instead of looking at the data as a line, they treat it like a 3D landscape that needs to be painted or "rendered."

Here is the breakdown using simple analogies:

1. The Old Way: The "Grid" Problem

Imagine you take a long strip of paper with your weather data written on it and fold it into a square grid to make it easier to look at.

• The Flaw: If you use standard image tools (like those used for photos) to look at this grid, the tool thinks the right edge of one row and the left edge of the next row are far apart.
• The Reality: In time, those two points are actually right next to each other (e.g., 11:59 PM and 12:00 AM).
• The Result: Old methods accidentally "cut" the timeline, creating a glitch where the computer thinks time jumps or breaks at the edges of the grid.

2. The New Idea: "Forecasting as Rendering"

The authors say: "Let's stop trying to predict a single number. Let's paint a picture of the future."

They use a technique borrowed from video games and 3D graphics called Gaussian Splatting.

• The Analogy: Imagine you have a bucket of glowing, fuzzy paint blobs (Gaussians). You don't just place one blob; you place thousands of them, overlapping each other, to create a smooth, continuous surface.
• The Magic: Because these blobs are "fuzzy" and overlap, they naturally connect the dots. If you place a blob at the end of a row, its "fuzziness" naturally spills over into the start of the next row, fixing the "cut" problem mentioned earlier.

3. How TimeGS Works (The Three Secret Ingredients)

A. The "Shape Shifter" (Multi-Basis Generation)

In the old 3D graphics world, you have to guess the exact shape and size of every single paint blob. This is hard and often leads to messy, unstable results.

• TimeGS Solution: Instead of guessing the shape from scratch, they have a pre-made library of shapes (a dictionary).
• The Analogy: Imagine you are an artist, but instead of mixing every color from scratch, you have a box of 50 perfect, pre-mixed paint tubes. You just decide how much of each tube to use to create your picture. This makes the painting process much faster, more stable, and less likely to make mistakes.

B. The "Seamless Stitch" (Chronologically Continuous Rasterization)

This is the part that fixes the "Grid Problem."

• The Analogy: Imagine you are tiling a floor, but the tiles are actually a continuous carpet. When you lay a carpet tile at the edge of the room, it doesn't stop; it wraps around and continues seamlessly on the other side.
• TimeGS Solution: The math ensures that the "fuzzy blobs" wrap around the edges of the time grid. This guarantees that the prediction for 11:59 PM flows perfectly into 12:00 AM without any breaks or glitches.

C. The "Smart Mixer" (Channel-Adaptive Aggregation)

Real-world data often has many different variables (e.g., temperature, humidity, wind speed). They all behave differently.

• The Analogy: Imagine you have a team of 5 different weather experts. One is great at predicting rain, another is great at wind, and another is great at temperature.
• TimeGS Solution: Instead of asking all experts for the same answer and averaging them, TimeGS has a "Smart Mixer." It listens to the specific needs of the day. If it's a windy day, it trusts the wind expert more. If it's a rainy day, it trusts the rain expert. It dynamically decides who to listen to for each specific variable.

Why is this a big deal?

• It's Smarter: By treating time as a continuous surface to be "rendered" rather than a list of numbers to be "calculated," it captures complex patterns (like sudden storms or weekly cycles) much better.
• It's Stable: Using the "pre-made shape library" prevents the computer from getting confused or making wild guesses.
• It Wins: In tests against the best existing AI models (like Transformers and other deep learning giants), TimeGS predicted the future more accurately across many different datasets (weather, electricity, traffic).

In short: TimeGS stops treating time like a broken line of numbers and starts treating it like a smooth, continuous painting, using a smart library of shapes and a seamless wrapping technique to predict the future with high precision.

Technical Summary

1. Problem Statement

Time Series Forecasting (TSF) faces significant challenges due to the complex entanglement of intra-period fluctuations (high-frequency variations within a cycle) and inter-period trends (long-term evolution across cycles).

Existing methods that attempt to model these patterns by reshaping 1D sequences into 2D period-phase representations (e.g., TimesNet, PDF) suffer from two critical limitations:

1. Topological Mismatch: Treating reshaped tensors as static images causes a break in chronological continuity at grid boundaries. Standard 2D spatial operators (like convolutions) treat the rightmost element of one row and the leftmost element of the next as spatially distant, whereas in time series, they are temporally adjacent ($t$ and $t+1$). This introduces artificial discontinuities.
2. Inefficient Uniform Representation: Fixed-size windowing allocates modeling capacity uniformly. However, real-world time series are often compressible; long stationary segments require few latent factors, while critical information lies in sparse change points or anomalies. Uniform representations fail to provide the adaptive resolution needed for these non-stationary patterns.

2. Methodology: TimeGS Framework

The authors propose TimeGS, a framework that shifts the forecasting paradigm from point-wise regression to 2D generative rendering. Instead of predicting discrete points, TimeGS reconstructs the future sequence as a continuous latent surface using 2D Gaussian Splatting (2DGS).

The architecture consists of four sequential components:

A. 2D Variation Feature Extraction (2D-VFE)

• Reshaping: Historical 1D sequences are folded into 2D tensors based on a defined period length $T$, creating a grid where rows represent periods and columns represent phases.
• Feature Encoding: Multiple parallel branches (UNet-based encoders) extract coupled variation features from diverse temporal views. This allows the model to capture both global inter-period trends and local intra-period dynamics simultaneously.

B. Multi-Basis Gaussian Kernel Generation (MB-GKG)

• The Challenge: Directly optimizing the position and covariance (shape) of free-floating Gaussian kernels on noisy time series data leads to training instability.
• The Solution: TimeGS reformulates shape regression as a stable dictionary learning task.
  • A Fixed Gaussian Basis Bank is pre-defined, containing a dictionary of Gaussian profiles with fixed positions and pre-computed covariance structures (Cholesky factors).
  • The model does not predict raw geometric parameters. Instead, it predicts mixing weights (via Softmax) and intensities (via MLP) to synthesize complex, anisotropic Gaussian kernels by linearly combining the fixed basis kernels. This ensures the generated shapes remain robust and plausible.

C. Multi-Period Chronologically Continuous Rasterization (MP-CCR)

• The Challenge: Standard 2D rasterization fails to handle the wrap-around nature of time series (where the end of one period connects to the start of the next).
• The Solution: The MP-CCR block treats Gaussian kernels as continuous signal segments.
  • It implements a rendering mechanism where the influence of a Gaussian kernel is calculated based on temporal distance rather than just 2D Euclidean distance.
  • The rendering process effectively "wraps" the kernel influence around the grid boundaries, ensuring that a Gaussian centered near the end of a period naturally influences the beginning of the next period, thereby enforcing strict chronological continuity.

D. Channel-Adaptive Aggregation (CCA)

• In multivariate forecasting, different variables may have distinct dominant periods or variation patterns.
• TimeGS employs learnable branch weights and component weights to adaptively fuse the outputs from multiple parallel branches. This allows the model to selectively leverage the most relevant "expert" view for each specific channel.

3. Key Contributions

1. Paradigm Shift: Pioneers the application of 2D Gaussian Splatting to time series forecasting, moving from regression to generative rendering. This explicitly models coupled temporal variations via anisotropic 2D Gaussians.
2. Topological Resolution: Introduces MP-CCR to resolve the boundary artifacts inherent in 2D reshaping, ensuring that the 2D representation respects the 1D sequential nature of time data.
3. Stable Optimization: Proposes MB-GKG, which replaces unstable geometric regression with a dictionary-based weight prediction task, stabilizing the training of Gaussian kernels on noisy data.
4. State-of-the-Art Performance: Demonstrates superior performance across diverse benchmarks, validating that structural rendering is more effective than traditional 1D or static 2D approaches for complex temporal dynamics.

4. Experimental Results

The authors evaluated TimeGS on 7 standard real-world datasets (including Weather, Electricity, Traffic, and ETT) against state-of-the-art baselines (Transformers, MLPs, and CNNs like TimesNet).

• Performance: TimeGS achieved State-of-the-Art (SOTA) results on most datasets and prediction horizons (96, 192, 336, 720 steps).
  • It significantly outperformed 1D models (e.g., iTransformer, PatchTST) and other 2D models (e.g., TimesNet).
  • Notable improvements were observed on datasets with strong periodicity and complex variations (e.g., Electricity and Traffic).
• Ablation Studies:
  • Removing MB-GKG: Caused catastrophic performance drops (e.g., MSE on Electricity surged from 0.181 to 0.877), proving the necessity of the fixed basis dictionary for stability.
  • Removing MP-CCR: Replacing the rendering mechanism with standard regression heads (Linear/MLP) resulted in inferior performance, confirming the value of the generative rendering approach.
  • Feature Visualization: t-SNE plots showed that different encoder branches learned distinct, non-redundant feature clusters, validating the multi-branch design.

5. Significance

This paper offers a fundamental rethinking of how time series data is modeled. By treating forecasting as a rendering problem, TimeGS bridges the gap between the structural benefits of 2D modeling and the sequential continuity of time series.

• Theoretical Insight: It demonstrates that the "topological mismatch" in existing 2D methods is a solvable problem through continuous rendering techniques rather than just better convolutional kernels.
• Practical Impact: The framework provides a robust mechanism for handling non-stationary, multi-scale temporal patterns, offering a new direction for deep learning in time series analysis beyond the Transformer and MLP paradigms.