Torch-Trade/btcusdt_perp_metrics_5m_09_2020_to_04_2026

BTCUSDT Perpetual Futures Metrics (5m)

Binance UM Futures metrics data for BTCUSDT, at native 5-minute cadence. Captures aggregate positioning and flow: open interest, top-trader long/short ratios, and taker buy/sell volume ratio.

Coverage

• Range: 2020-09-01 00:00 → 2026-04-24 00:00 (UTC)
• Cadence: 5 minutes (288 rows/day nominal)
• Rows: 592,934
• Completeness vs 5-min grid: 99.893%
• Missing days (no file published): 0

Columns

Column	Type	Description
timestamp	datetime64[ns]	Metric sample time (UTC), 5-min grid
sum_open_interest	float64	Total open interest in BTC
sum_open_interest_value	float64	Total open interest in USD
count_toptrader_long_short_ratio	float64	Top-trader account L/S ratio (N_long / N_short accounts)
sum_toptrader_long_short_ratio	float64	Top-trader position L/S ratio (position value long / short)
count_long_short_ratio	float64	All-account L/S ratio
sum_taker_long_short_vol_ratio	float64	5-min taker buy / sell volume ratio

Meaning:

• sum_taker_long_short_vol_ratio > 1 in a 5-min window = more market buys than sells = upward pressure.
• sum_toptrader_long_short_ratio > 1 = top traders net-long. Watch for extreme values (>2 or <0.5) as positioning stress.
• sum_open_interest_value delta reveals accumulation (OI up + price up = new longs) vs distribution (OI down + price up = shorts covering).

Source

Binance publishes these files daily at:

https://data.binance.vision/data/futures/um/daily/metrics/BTCUSDT/BTCUSDT-metrics-YYYY-MM-DD.zip

This dataset concatenates all available daily files in the covered range, deduplicates on timestamp, and saves as a single parquet.

Usage

from datasets import load_dataset
import pandas as pd

ds = load_dataset("Torch-Trade/btcusdt_perp_metrics_5m_09_2020_to_04_2026", split="train")
metrics = ds.to_pandas().sort_values("timestamp").reset_index(drop=True)
print(metrics.head())

Joining with 1-minute OHLCV

Metrics are 5-min native. To join with 1-min bars, forward-fill onto the 1-min grid:

import pandas as pd
from datasets import load_dataset

ohlcv = load_dataset("Torch-Trade/btcusdt_perp_1m_05_2021_to_02_2026", split="train").to_pandas()
metrics = load_dataset("Torch-Trade/btcusdt_perp_metrics_5m_09_2020_to_04_2026", split="train").to_pandas()

ohlcv["timestamp"] = pd.to_datetime(ohlcv["timestamp"])
metrics["timestamp"] = pd.to_datetime(metrics["timestamp"])

joined = pd.merge_asof(
    ohlcv.sort_values("timestamp"),
    metrics.sort_values("timestamp"),
    on="timestamp",
    direction="backward",   # forward-fill: each 1-min bar gets the most recent metric
)

Derived features

Common transforms at training time:

• oi_change = sum_open_interest.pct_change(12) — 1-hour OI delta at 5-min cadence
• ls_ratio_zscore = (sum_toptrader_long_short_ratio - rolling_mean_24h) / rolling_std_24h
• taker_flow_ema = sum_taker_long_short_vol_ratio.ewm(span=12).mean()

License

Data © Binance. Redistributed under fair use for research purposes. See https://data.binance.vision/ for upstream licensing.