# Usage guide

## Understand what each series means

- Read `DATA_DICTIONARY.md` for plain-language definitions grouped by category.
- Read `DATA_CATALOG.md` for endpoint, unit, quality notes, and materialization status.
- Download from the `dataset-latest` GitHub Release for the complete generated dataset. A Git clone contains collector code, configuration, and docs, not the generated dataset.

## Download the parquet training archive (recommended)

This archive is the recommended way for model training.
It includes only parquet series plus documentation files.

```bash
curl -L -o resonance-data-train-parquet.zip \
  https://github.com/JrVillabona/resonance-data/releases/download/dataset-latest/resonance-data-train-parquet.zip
unzip -o resonance-data-train-parquet.zip
```

Optional integrity verification:

```bash
curl -L -o resonance-data-train-parquet.zip.sha256 \
  https://github.com/JrVillabona/resonance-data/releases/download/dataset-latest/resonance-data-train-parquet.zip.sha256
shasum -a 256 -c resonance-data-train-parquet.zip.sha256
```

Optional manifest inspection:

```bash
curl -L -o resonance-data-train-parquet.manifest.json \
  https://github.com/JrVillabona/resonance-data/releases/download/dataset-latest/resonance-data-train-parquet.manifest.json
cat resonance-data-train-parquet.manifest.json
```

## Download the full archive (csv + parquet + meta)

Use this only if you need raw csv files or metadata sidecars.

```bash
curl -L -o resonance-data-latest.zip \
  https://github.com/JrVillabona/resonance-data/releases/download/dataset-latest/resonance-data-latest.zip
unzip -o resonance-data-latest.zip
```

The full archive also publishes `resonance-data-latest.manifest.json` and `resonance-data-latest.zip.sha256`.

## Load from Hugging Face

Researchers can load the full long-format Parquet table directly from Hugging Face:

```python
from datasets import load_dataset

ds = load_dataset("JrVillabona/resonance-data", split="train")
```

Or with Pandas/Polars through the Hub filesystem:

```python
import pandas as pd

df = pd.read_parquet("hf://datasets/JrVillabona/resonance-data/data/train.parquet")
```

Direct browser or CLI download:

- `https://huggingface.co/datasets/JrVillabona/resonance-data`
- `https://huggingface.co/datasets/JrVillabona/resonance-data/resolve/main/data/train.parquet`

```bash
curl -L -o resonance-train.parquet \
  https://huggingface.co/datasets/JrVillabona/resonance-data/resolve/main/data/train.parquet
```

Long-format columns:

- `symbol`
- `date`
- `open`, `high`, `low`, `close`, `volume`
- `source`, `ticker`, `category`, `data_type`, `frequency`, `description`

## Load one series

```python
import pandas as pd

df = pd.read_parquet("data/BTC_daily.parquet")
```

## Verify CSV and Parquet parity

```python
import pandas as pd

csv_df = pd.read_csv("data/BTC_daily.csv", parse_dates=["date"])
pq_df = pd.read_parquet("data/BTC_daily.parquet")
assert len(csv_df) == len(pq_df)
assert csv_df["date"].min() == pq_df["date"].min()
assert csv_df["date"].max() == pq_df["date"].max()
```

## Align mixed frequencies

- Daily: crypto prices, yields, commodities
- Weekly: selected FRED series
- Monthly: CPI, M2, labor metrics

Recommended merge strategy:

1. Build a daily index from min/max dates.
2. Left-join all series.
3. Forward-fill lower-frequency series after release date.
4. Keep publication lags in feature engineering logic.