afeng commited on
Commit
5e30441
·
verified ·
1 Parent(s): a3da3b5

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +110 -0
README.md CHANGED
@@ -29,3 +29,113 @@ configs:
29
  - split: train
30
  path: data/train-*
31
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29
  - split: train
30
  path: data/train-*
31
  ---
32
+ # MTBench: A Multimodal Time Series Benchmark
33
+
34
+
35
+ **MTBench** ([Huggingface](https://huggingface.co/collections/afeng/mtbench-682577471b93095c0613bbaa), [Github](https://github.com/Graph-and-Geometric-Learning/MTBench), [Arxiv](https://arxiv.org/pdf/2503.16858)) is a suite of multimodal datasets for evaluating large language models (LLMs) in temporal and cross-modal reasoning tasks across **finance** and **weather** domains.
36
+
37
+ Each benchmark instance aligns high-resolution time series (e.g., stock prices, weather data) with textual context (e.g., news articles, QA prompts), enabling research into temporally grounded and multimodal understanding.
38
+
39
+ ## 🏦 Stock Time-Series and News Pair
40
+
41
+ This dataset contains aligned pairs of financial news articles and corresponding stock time-series data, designed to evaluate models on **event-driven financial reasoning** and **news-aware forecasting**.
42
+
43
+ ### Pairing Process
44
+
45
+ Each pair is formed by matching a news article’s **publication timestamp** with a relevant stock’s **time-series window** surrounding the event
46
+
47
+ To assess the impact of the news, we compute the **average percentage price change** across input/output windows and label directional trends (e.g., `+2% ~ +4%`). A **semantic analysis** of the article is used to annotate the sentiment and topic, allowing us to compare narrative signals with actual market movement.
48
+
49
+ We observed that not all financial news accurately predicts future price direction. To quantify this, we annotate **alignment quality**, indicating whether the sentiment in the article **aligns with observed price trends**. Approximately **80% of the pairs** in the dataset show consistent alignment between news sentiment and trend direction.
50
+
51
+
52
+ ### Each pair includes:
53
+
54
+ - `"input_timestamps"` / `"output_timestamps"`: Aligned time ranges (5-minute resolution)
55
+ - `"input_window"` / `"output_window"`: Time-series data (OHLC, volume, VWAP, transactions)
56
+ - `"text"`: Article metadata
57
+ - `content`, `timestamp_ms`, `published_utc`, `article_url`
58
+ - Annotated `label_type`, `label_time`, `label_sentiment`
59
+ - `"trend"`: Ground truth price trend and bin labels
60
+ - Percentage changes and directional bins (e.g., `"-2% ~ +2%"`)
61
+ - `"technical"`: Computed technical indicators
62
+ - SMA, EMA, MACD, Bollinger Bands (for input, output, and overall windows)
63
+ - `"alignment"`: Label indicating semantic-trend consistency (e.g., `"consistent"`)
64
+
65
+
66
+
67
+ ## 📦 Other MTBench Datasets
68
+
69
+ ### 🔹 Finance Domain
70
+
71
+ - [`MTBench_finance_news`](https://huggingface.co/datasets/afeng/MTBench_finance_news)
72
+ 20,000 articles with URL, timestamp, context, and labels
73
+
74
+ - [`MTBench_finance_stock`](https://huggingface.co/datasets/afeng/MTBench_finance_stock)
75
+ Time series of 2,993 stocks (2013–2023)
76
+
77
+ - [`MTBench_finance_aligned_pairs_short`](https://huggingface.co/datasets/afeng/MTBench_finance_aligned_pairs_short)
78
+ 2,000 news–series pairs
79
+ - Input: 7 days @ 5-min
80
+ - Output: 1 day @ 5-min
81
+
82
+ - [`MTBench_finance_aligned_pairs_long`](https://huggingface.co/datasets/afeng/MTBench_finance_aligned_pairs_long)
83
+ 2,000 news–series pairs
84
+ - Input: 30 days @ 1-hour
85
+ - Output: 7 days @ 1-hour
86
+
87
+ - [`MTBench_finance_QA_short`](https://huggingface.co/datasets/afeng/MTBench_finance_QA_short)
88
+ 490 multiple-choice QA pairs
89
+ - Input: 7 days @ 5-min
90
+ - Output: 1 day @ 5-min
91
+
92
+ - [`MTBench_finance_QA_long`](https://huggingface.co/datasets/afeng/MTBench_finance_QA_long)
93
+ 490 multiple-choice QA pairs
94
+ - Input: 30 days @ 1-hour
95
+ - Output: 7 days @ 1-hour
96
+
97
+ ### 🔹 Weather Domain
98
+
99
+ - [`MTBench_weather_news`](https://huggingface.co/datasets/afeng/MTBench_weather_news)
100
+ Regional weather event descriptions
101
+
102
+ - [`MTBench_weather_temperature`](https://huggingface.co/datasets/afeng/MTBench_weather_temperature)
103
+ Meteorological time series from 50 U.S. stations
104
+
105
+ - [`MTBench_weather_aligned_pairs_short`](https://huggingface.co/datasets/afeng/MTBench_weather_aligned_pairs_short)
106
+ Short-range aligned weather text–series pairs
107
+
108
+ - [`MTBench_weather_aligned_pairs_long`](https://huggingface.co/datasets/afeng/MTBench_weather_aligned_pairs_long)
109
+ Long-range aligned weather text–series pairs
110
+
111
+ - [`MTBench_weather_QA_short`](https://huggingface.co/datasets/afeng/MTBench_weather_QA_short)
112
+ Short-horizon QA with aligned weather data
113
+
114
+ - [`MTBench_weather_QA_long`](https://huggingface.co/datasets/afeng/MTBench_weather_QA_long)
115
+ Long-horizon QA for temporal and contextual reasoning
116
+
117
+
118
+
119
+ ## 🧠 Supported Tasks
120
+
121
+ MTBench supports a wide range of multimodal and temporal reasoning tasks, including:
122
+
123
+ - 📈 **News-aware time series forecasting**
124
+ - 📊 **Event-driven trend analysis**
125
+ - ❓ **Multimodal question answering (QA)**
126
+ - 🔄 **Text-to-series correlation analysis**
127
+ - 🧩 **Causal inference in financial and meteorological systems**
128
+
129
+
130
+
131
+ ## 📄 Citation
132
+
133
+ If you use MTBench in your work, please cite:
134
+
135
+ ```bibtex
136
+ @article{chen2025mtbench,
137
+ title={MTBench: A Multimodal Time Series Benchmark for Temporal Reasoning and Question Answering},
138
+ author={Chen, Jialin and Feng, Aosong and Zhao, Ziyu and Garza, Juan and Nurbek, Gaukhar and Qin, Cheng and Maatouk, Ali and Tassiulas, Leandros and Gao, Yifeng and Ying, Rex},
139
+ journal={arXiv preprint arXiv:2503.16858},
140
+ year={2025}
141
+ }