Technical Summary: Timeflies – Existence Precedes Value

1. Problem Statement

Real-world time series forecasting is frequently hindered by high levels of incompleteness and irregularity caused by sensor dormancy, transmission delays, and event-driven sampling. While existing methods have evolved from rigid "impute-then-forecast" pipelines to continuous-time models (e.g., Neural ODEs, continuous-time graph networks) that better handle historical irregularities, they suffer from a critical limitation: the perfect foresight assumption.

Current state-of-the-art models assume that the timestamps of future valid observations are known in advance during inference. In practical systems, however, it is often unclear whether a valid observation will occur at a specific future moment. Predicting a numerical value without first confirming its existence is often impractical, as decision-making processes rely on the observability of information. The paper argues that missingness should not be treated merely as a static data defect to be imputed, but as a profound temporal signal reflecting the system's behavioral patterns.

2. Methodology: The Timeflies Framework

To address this, the authors propose Timeflies, a unified forecasting framework that reformulates time series prediction as a joint problem of future observability inference and value estimation. The core philosophy is "Existence Precedes Value": a numerical prediction is only meaningful if the observation is actually realized.

Architecture Overview

Timeflies employs a symmetric dual-stream architecture consisting of an Observation Stream and a Value Stream, coupled through three synergistic modules:

1. Reliability-Aware Patch Embedding:

   • The input sequence is divided into patches.
   • A Mask-Aware Normalization is applied to the value sequence, computing statistics only on observed entries to prevent distortion from missing values.
   • The observation stream encodes missing positions and intervals using a gated fusion mechanism.
   • A Patch Reliability Score ($\rho$) is computed based on the observation ratio and missing interval duration within each patch. This score modulates the injection of observation patterns into the value embedding, attenuating contributions from sparse, noise-dominated patches while enhancing reliable ones.

2. Observation-Guided Value Attention:

   • Both streams utilize self-attention to capture long-range dependencies.
   • Crucially, the observation stream does not operate in isolation; it guides the value stream via a Cross-track Conditioned Attention mechanism.
   • The attention map is a fusion of the value stream's semantic correlations, the observation stream's structural correlations, and a reliability-weighted term. This ensures that value predictions are informed by historical missingness patterns, effectively routing attention away from unreliable sparse keys and toward informative patches.

3. Dual Prediction Head:

   • The high-level representations from both streams are mapped to the prediction horizon.
   • Observability Prediction: A classification head predicts the probability of a valid observation occurring at each future time step ($\hat{O}$).
   • Value Prediction: A regression head predicts the future state values ($\hat{Y}$).
   • Gated Fusion: The predicted observability acts as a soft gate on the value prediction. The final value estimate is a combination of the direct value stream output and the observation stream's contribution, weighted by the predicted existence probability. This ensures that observation features significantly influence the prediction only when existence is highly probable.

Training Objective

The model is trained with a joint loss function:
$$\mathcal{L} = \mathcal{L}_{val} + \eta \cdot \mathcal{L}_{obs}$$

• $\mathcal{L}_{val}$: Mean Squared Error (MSE) calculated only on time steps where ground-truth observations exist.
• $\mathcal{L}_{obs}$: Focal Binary Cross-Entropy (Focal BCE) to handle the class imbalance between observed and missing positions.
• $\eta$: A hyperparameter balancing the two tasks.

3. Key Contributions

A New Forecasting Paradigm

The paper challenges the "oracle timestamps assumption" prevalent in previous models. It pioneers a paradigm where the model explicitly predicts whether an observation will occur before estimating what its value will be. This shifts the focus from simple numerical extrapolation to inferring the information generation mechanism.

Technical Advancement (Timeflies)

Timeflies explicitly decomposes time series into two interactive streams. By extracting historical observation rhythms as dynamic context, the architecture enables existence patterns to guide and refine state predictions. This deep interaction bridges the gap between the observability of information and its underlying numerical values.

Empirical and Practical Utility

• Shadow Benchmark: To rigorously validate the framework, the authors constructed Shadow, a comprehensive benchmark combining natural missing gaps from public datasets (GIFT-Eval) with real-world industrial e-commerce data. Crucially, the proprietary datasets contain strictly natural missing values (non-random, encoding structural patterns like off-peak inactivity) with missing ratios ranging from <0.1% to >89%.
• OVJE Metric: The authors introduce the Observation-Value Joint Entropy (OVJE) metric. Unlike traditional marginal error metrics (MSE/MAE), OVJE evaluates the coupled capability of existence inference and value estimation by computing the negative log-likelihood of the joint distribution $P(y_t, o_t | x_t)$.

4. Experimental Results

Extensive experiments on the Shadow benchmark demonstrate that Timeflies consistently outperforms existing baselines, including Transformer-based (PatchTST, iTransformer, FEDformer), Linear-based (OLinear, DLinear), and CNN-based methods.

• Performance: Timeflies achieves state-of-the-art performance across all sparsity regimes (No, Low, Medium, High missingness).
• Sparsity Sensitivity: The performance gap widens significantly as data sparsity increases. In high-sparsity environments, traditional baselines degrade by blindly extrapolating values, whereas Timeflies effectively isolates true signals from missing-induced noise.
• Robustness: On datasets with extreme missingness shifts (e.g., training with <1% missingness and testing with >90%), Timeflies maintains robust performance, while competing models fail catastrophically.
• Ablation Studies: Removing any core component (Mask-Aware Normalization, Observation-Guided Attention, Patch Reliability, or Focal Loss) leads to performance degradation, confirming the structural necessity of each module.

5. Significance and Claims

The paper claims that explicitly modeling future observability is an indispensable gate for reliable numerical prediction in irregular settings. By treating missingness as a structured prior rather than a defect, Timeflies enables decision-making systems to incorporate observation uncertainty.

The authors modestly note that their work focuses on the GIFT-Eval benchmark with naturally occurring missing patterns and assumes a stationary missingness mechanism. They acknowledge limitations such as the constant computational overhead of observation-aware modules even in fully observed settings and the assumption of channel independence. However, the work opens a new avenue for time series forecasting by bridging observability and numerical regression, offering a more realistic approach for real-world deployments where data integrity cannot be guaranteed.