Bridging the High-Frequency Data Gap: A Millisecond-Resolution Network Dataset for Advancing Time Series Foundation Models

This paper introduces a novel millisecond-resolution dataset derived from operational 5G wireless and traffic conditions to address the high-frequency data gap in time series foundation models, demonstrating that current models struggle with such data and highlighting the critical need for high-frequency pre-training to improve their real-world robustness and generalization.

The Gist

Imagine you are trying to teach a super-smart robot how to predict the future. You want this robot to be good at guessing everything from the weather next week to the stock market tomorrow. To do this, you feed it a massive library of history books (datasets) so it can learn patterns.

For a long time, these history books were written in slow motion. They recorded events like "temperature at 12:00 PM" or "stock price at the end of the day." The robot learned to predict things that changed slowly, like the seasons or the economy.

The Problem: The Robot is Blind to the Blink of an Eye
The authors of this paper realized there was a huge gap. Real life isn't always slow. In the world of 5G wireless networks, things change in the blink of an eye—literally milliseconds. Traffic jams, video buffering, or a sudden cyber-attack happen so fast that the "slow-motion" robot is completely lost. It's like trying to catch a hummingbird with a butterfly net designed for slow-moving moths.

Current "Time Series Foundation Models" (the fancy name for these super-robots) are trained on slow data. When you throw them into a fast-paced 5G network, they stumble. They can't see the tiny, rapid spikes and drops that define how these networks actually work.

The Solution: A New "High-Speed" Training Manual
The team created a brand-new dataset. Think of this as a high-definition, high-speed camera recording of a 5G network.

• The Source: They set up a real 5G network in a lab (the OpenIreland testbed).
• The Action: They simulated people walking, driving, and even hackers trying to crash the system (DDoS attacks).
• The Resolution: Instead of recording once a second, they recorded 10 times every second (every 100 milliseconds). This captures the "heartbeat" of the network in real-time.

The Experiment: The Race
To test if their new "high-speed manual" helps, they pitted two types of learners against each other:

1. The Old Guard (Shallow Models): These are like experienced, old-school mechanics. They don't have a massive library of knowledge, but they are very good at looking at the immediate past and adjusting quickly.
2. The New Guard (Foundation Models): These are the super-robots with massive libraries of slow-motion data. They are powerful but rigid.

The Results: The Underdog Wins
Here is the twist: The old-school mechanics (specifically a model called Adaptive Random Forest) beat the super-robots.

• Why the Super-Robots Failed: The robots were trained on smooth, predictable data (like electricity usage or weather). When they saw the 5G network, it looked like chaos—sudden spikes, weird jumps, and noise. The robots tried to apply their "slow-motion rules" to a "fast-motion world" and failed. Even when they tried to "fine-tune" (re-learn) on the new data, they struggled to adapt.
• Why the Mechanics Won: The old-school models are designed to adapt on the fly. They don't rely on long-term patterns; they react to what just happened right now. In a chaotic, fast-changing environment, this flexibility is king.

The Big Takeaway
The paper isn't saying the super-robots are useless. It's saying we are missing a crucial piece of the puzzle.

If we want these AI models to be truly smart and useful in the real world (like managing 6G networks, autonomous cars, or high-frequency trading), we can't just feed them slow data. We need to feed them fast data too.

The Analogy:
Imagine you are training a pilot.

• Current Method: You only let them practice in a simulator where the plane flies in a straight line at 100 mph.
• The Reality: They will eventually have to fly a fighter jet in a storm at 1,000 mph with sudden turbulence.
• The Paper's Point: If you only train them on the slow simulator, they will crash when they hit the storm. You need to give them a simulator that includes the storm, the turbulence, and the speed.

In Summary:
This paper introduces a new, ultra-fast dataset for 5G networks. It proves that current AI models are too "slow" to handle real-world, high-speed data. To build better AI for the future, we need to stop training them only on slow, calm data and start teaching them how to dance to the rhythm of the fast, chaotic, real world.

Technical Summary

1. Problem Statement

Time Series Foundation Models (TSFMs) have demonstrated significant success in generalizing across various domains (e.g., energy, finance, weather) and tasks. However, a critical limitation exists: current large-scale pre-training datasets predominantly consist of low-frequency time series (sampling intervals ranging from seconds to years).

• The Gap: This lack of high-frequency data (millisecond resolution) hinders TSFMs from capturing the nuances of real-world systems that operate at rapid speeds, such as wireless networks.
• The Consequence: Existing TSFMs struggle to generalize to high-frequency domains, exhibiting poor performance in zero-shot and fine-tuned settings when applied to data characterized by rapid fluctuations, non-stationarity, and irregular spikes.

2. Methodology

A. Dataset Construction (The Core Contribution)

The authors introduce a novel dataset derived from a real-world 5G Open Radio Access Network (O-RAN) deployment within the OpenIreland testbed.

• Source: Data was collected using Software-Defined Radios (Ettus USRPs) acting as a base station and multiple User Equipments (UEs).
• Resolution: Millisecond-level sampling, aggregated into 100-millisecond intervals to balance granularity with O-RAN operational overhead.
• Scope: The dataset covers diverse mobility patterns (static, pedestrian, car, bus, train) and traffic types (benign: web, VoIP, video; malicious: DDoS, PortScan, etc.).
• Target Variable: Downlink Bitrate (mac_dl_brate), with multivariate inputs including Channel Quality Indicator (CQI), Modulation and Coding Scheme (MCS), packet counts, and signal strength.
• Characteristics: Unlike standard datasets (e.g., ETTh1, Electricity), this data is non-stationary and heteroskedastic. It features:
  • Unstable, step-like trends.
  • Weak short-term seasonality (often obscured by noise).
  • Heavy-tailed residual distributions with sharp spikes and bursts of volatility.
  • Strong temporal persistence with slow decay in autocorrelation.

B. Benchmarking Framework

The authors conducted a comprehensive evaluation comparing Shallow Machine Learning Models against State-of-the-Art TSFMs.

• Shallow Models:
  • Random Forest (RF), XGBoost (XGB), Online Linear Regression (OLR), and Naive Forecast.
  • Adaptive Random Forest (ARF): Specifically highlighted for its ability to handle concept drift in data streams.
• TSFMs:
  • TinyTimeMixer (TTM): A lightweight pre-trained model.
  • Chronos: A language-modeling-based probabilistic TSFM (evaluated in zero-shot and fine-tuned modes).
  • Lag-Llama: A decoder-only transformer-based foundation model.
• Experimental Setup:
  • Task: Short-term forecasting with horizons from 100ms (1 step) to 9.6 seconds (96 steps).
  • Settings: Both Univariate and Multivariate (using 4 to 10 features).
  • Evaluation Metrics: Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).
  • Protocol: 80/20 train/test split preserving temporal order; rolling evaluation for TSFMs to match the sample-level evaluation of shallow models.

3. Key Contributions

1. Novel High-Frequency Dataset: Introduction of the first millisecond-resolution wireless network dataset, filling a critical gap in the TSFM pre-training ecosystem. It expands the domain scope beyond energy and finance into telecommunications.
2. Domain Characterization: Detailed statistical analysis revealing that wireless network data differs fundamentally from existing benchmarks (unstable trends, heavy-tailed noise, lack of smooth seasonality).
3. Comprehensive Benchmark: A rigorous evaluation demonstrating that current TSFMs fail to generalize to high-frequency, non-stationary network data, while adaptive shallow models (specifically ARF) excel.
4. Ablation Studies: Analysis of fine-tuning strategies (Head-only vs. Adapter-based) and the impact of temporal resolution, showing that simply increasing resolution does not fix TSFM performance issues; the data distribution itself is the bottleneck.

4. Key Results

• Performance Gap:
  • Adaptive Random Forest (ARF) consistently outperformed all other models in both univariate and multivariate settings.
  • TSFMs (TTM, Chronos, Lag-Llama) performed poorly, often worse than or comparable to simple baselines like Naive forecasting, even after fine-tuning.
  • Zero-Shot vs. Fine-Tuning: While fine-tuning TSFMs on this dataset improved their performance slightly compared to zero-shot, they still failed to match the accuracy of ARF.
• Quantitative Evidence (Multivariate RMSE):
  • ARF: ~0.0175 (Best)
  • Chronos (Fine-tuned): ~0.0253
  • TTM (Fine-tuned): ~0.0393
  • Naive: ~0.0418
• Visual Analysis: Predictions from TSFMs failed to track the rapid, step-like changes and spikes in the actual bitrate, often smoothing over critical dynamics. ARF successfully adapted to these abrupt shifts.
• Ablation Findings:
  • Fine-tuning Strategies: Head-only and Adapter-based fine-tuning for TTM yielded worse results than the default fine-tuning approach.
  • Temporal Resolution: Increasing the temporal resolution (e.g., from 100ms to 2000ms) did not significantly improve TSFM performance, confirming that the issue is the inherent data distribution (spikes, non-stationarity) rather than just the frequency.

5. Significance and Future Implications

• Redefining Pre-training Requirements: The paper argues that for TSFMs to be robust in real-world applications (like 5G/6G network control), pre-training datasets must include high-frequency, non-stationary, and irregular data. Current models are over-optimized for smooth, low-frequency trends.
• Architectural Limitations: The results suggest that current Transformer-based architectures may lack the inductive biases necessary to model rapid concept drift and heavy-tailed noise without specific architectural adaptations or exposure to similar data during pre-training.
• Practical Applications: The dataset enables new use cases in O-RAN, including:
  • Predictive Control: Proactive handover and resource allocation based on millisecond-level throughput predictions.
  • Anomaly Detection: Identifying DDoS attacks or congestion via deviations in CQI and SINR.
  • Mobility Management: Classifying user movement (e.g., train vs. pedestrian) to optimize network slicing.
• Future Work: The authors propose using this dataset to develop specialized pre-training strategies, transfer learning across mobility profiles, and advanced anomaly detection mechanisms that do not rely on deep packet inspection.

In conclusion, this work serves as a critical "reality check" for the TSFM community, demonstrating that without high-frequency, domain-specific data, foundation models cannot achieve the generalization required for next-generation wireless network management.