mrzdu/oil_WTI_image_dataset_v1

FINANCE_PROJECT – Plot Image Dataset (Parquet Shards)

This dataset contains Matplotlib-rendered PNG plot images generated from sliding windows over financial time series features. Each row is one image sample plus 1–8 day trend labels derived from CL=F_Close, where labels are defined at the last index of the plotted window.

Repository Layout

The dataset is stored as Hugging Face–style sharded Parquet files:

• train-00000-of-00210.parquet
• train-00001-of-00210.parquet
• …
• train-00209-of-00210.parquet

This is the intended structure. Avoid mixing in raw builder outputs like part_00000x.parquet in the same repo, because it makes the dataset view confusing and inconsistent.

Data Sources (Yahoo Finance)

The dataset is built from:

• ^OVX (CBOE Crude Oil Volatility Index) → Close price
• CL=F (WTI Crude Oil Futures) → OHLCV and derived features

Typical generation settings:

• Frequency: Daily (D)
• Date range (example): 2000-01-01 to 2025-01-01
• Window length: 50

What Each Row Represents

Each row corresponds to:

1. A sliding window of length window over the dataframe

   • end_idx = start_idx + window - 1

2. A plot image generated from a selected feature group

   • Dark (black) background
   • Bright, deterministic feature colors

3. Trend labels computed from the window end (end_idx)

   • Labels belong to the last index of the plotted window, not the first.

Label Definition (1–8 Day Trend)

Labels are computed using only CL=F_Close.

For each horizon h in {1,2,3,4,5,6,7,8}:

• pct = Close[end_idx + h] / Close[end_idx] - 1

Ternary Mode (UP / FLAT / DOWN)

With pct_threshold = 0.002 (0.2%):

• +1 if pct > +0.002 → UP
• 0 if |pct| <= 0.002 → FLAT
• -1 if pct < -0.002 → DOWN

Typical label columns:

• trend_1d, trend_2d, …, trend_8d

Note: Windows near the end of the series are skipped if future values (end_idx + h) do not exist.

Columns (Typical Parquet Schema)

Each row typically includes:

• image_png (bytes): PNG bytes of the plot image.
• plot_kind (string): Example "line" (or "candle" if candlesticks are used).
• start_idx (int): Window start index.
• end_idx (int): Window end index (labels are defined here).
• start_ts / end_ts (timestamp-like): Start/end timestamps if the index is datetime.
• columns_json (string): JSON list of plotted feature columns.
• trend_1d … trend_8d (int): Trend labels (ternary mode: -1, 0, +1).

Exact columns may vary slightly depending on plotting configuration.

Feature Groups

Each image is generated from a feature group consisting of:

• A constant column from the CL=F prefix (commonly CL=F_Close)
• Plus additional columns selected from the dataset to reach a fixed group size (e.g., k_total=4)

Example (k_total=4):

• [CL=F_Close, ^OVX_Close, CL=F_sma_20, CL=F_rsi_14]

The exact plotted feature list for each sample is stored in columns_json.

Color Mapping

Plots use a black background and bright colors.

Feature colors are assigned deterministically so that the same feature name (e.g., sma_20) maps to the same color across different plots and runs.

Efficient Reading Notes

Because image_png contains raw PNG bytes, loading the entire dataset into a single pandas DataFrame can consume a lot of RAM. For training and inspection:

• Prefer streaming / shard-by-shard reading
• Read only required columns (e.g., image_png + labels)

Disclaimer

This dataset is derived from market data accessed via Yahoo Finance. Ensure your usage complies with applicable terms and licensing requirements.

---