Kestrel: crypto signals, simulated trades and order-book snapshots

Records from Kestrel, a real-time signal detection and execution daemon that ran from June to November 2026 on 34 liquid USDT pairs. Over that time about 1,600 strategy instances generated signals on 5-minute, 1-hour and 4-hour candles and executed them with a simulated (paper) perpetual-futures execution engine that models fees, slippage, isolated margin and liquidation. No real money was traded.

Headline result: no edge after fees

Simulated trades	167,164 (166,991 closed)
Win rate	55.2%
PnL before fees	−$949
Fees + slippage	$1,801
Net PnL	−$2,748

Positions were risk-sized out of $10 isolated-margin buckets: median margin $1, about $20 notional, at 20× leverage (one cohort used 3×). Most fills were charged maker fees, a median round trip of 0.04% of notional. Every timeframe and every research tier was net-negative. This is published as a negative result and as raw data for anyone who wants to test their own ideas against it.

Tables

Config	Rows	What it is
trades (default)	167,164	Every simulated trade: entry/exit time and price, take-profit, stop-loss, liquidation price, size, leverage, fees, gross/net PnL, close reason
signals	68,472	Every signal evaluated from 2026-10-14 to 2026-11-02, fired or rejected, with the filter layers it passed and the rejection reason
candles	3.78M	5m / 1h / 4h OHLCV candles for 34 pairs (2026-08-04 to 2026-11-02) with the indicators computed at close: EMA 9/21, RSI 14, ATR 14, Bollinger bands, ADX, volume ratio, regime label, candle geometry
trade_context	55.0M	Links each trade to the candles in the 48 h before entry (pre), while open (during) and 48 h after exit (post)
microstructure	5.53M	Order-book and trade-tape snapshots every ~12.5 s for BTC, ETH, SOL, XRP, DOGE and ADA (2026-06-21 to 2026-11-02): best bid/ask, spread, depth at 5 and 20 levels, depth imbalance, aggressor buy/sell volume
pattern_memory	274	Running win statistics per pattern × direction × trading session × market regime

Conventions and caveats

• Timestamps are Unix milliseconds, UTC (ts, entry_ts, exit_ts, candle_ts).
• env is the research tier: dev (broad data collection), lab (curated) and staging (pre-production). All three used the same simulated execution engine.
• trades.source says where each row came from. The live trade table was wiped at several research resets, so earlier trades were recovered from database snapshots taken before each reset (kestrel-lean-…, kestrel-prereset-…, kestrel-full-…) and merged, with duplicates removed. Slates that were reset without a snapshot are missing, so coverage before October has gaps.
• IDs are only unique within one source. trade_context.trade_id and signals.trade_id join to trades.id only for rows where source = 'live_db_2026-11-02'.
• Candles are stored once per strategy instance (bot_id), so the same pair/timeframe/timestamp appears many times. Deduplicate on (pair, timeframe, ts) if you only need prices.
• Fill model. Taker: 0.04% fee + 0.05% slippage per side (~0.18% round trip). Maker (post-only entries and take-profits): 0.02% per side. Take-profit and stop-loss were checked on candle close. Liquidation price for longs is entry × (1 − 1/leverage + 0.005), and the mirror image for shorts. Perpetual funding was not charged in dev and was charged at 0.01% per 8 h in staging.
• close_reason values: take_profit, stop_loss, timeout (maximum hold reached), trailing_stop, signal_exit / indicator_tp (exit on an opposite signal), liquidated, manual, and orphaned_crash_recovery (closed at roughly zero PnL when the daemon restarted).
• Market data comes from Gate.io public feeds. Exchange terms may apply to redistributing raw market data.

What was tested

Each item below was tested both live (simulated) and in walk-forward backtests on two separate years:

• Candle-pattern entries on 5-minute candles (momentum, breakout, mean-reversion, wick rejection, engulfing, fair-value-gap retest, turtle soup, MACD / CCI / SMA crosses, cross-pair reversal, and others). Before fees they average about zero, and their win rate matches random entries with the same exits.
• High win-rate brackets (take-profit 0.5 × ATR, stop-loss 1.5 × ATR) reach about 70% wins but still lose: break-even at that geometry needs about 83%.
• Maker (post-only) execution roughly halves the loss but never makes it positive.
• Inverting every signal also loses, which shows the loss is cost rather than a direction that can be flipped.
• Time-of-day filters looked strong live but failed on the prior year.
• Order-book imbalance predicts the next move by only ~0.4–0.7 bps against a ~4 bps cost floor.
• Funding-rate carry (long spot, short perpetual) yields a low single-digit percentage per year and turns negative in flat or bear years.
• Slower cohorts (4-hour trend following and daily cross-pair momentum) were the only designs that cleared costs in both backtest years. They were still being evaluated live when recording stopped, and they were slightly negative at that point.

Loading

from datasets import load_dataset

trades = load_dataset("Azzindani/kestrel", "trades", split="train").to_pandas()
closed = trades[trades.exit_ts.notna()]
print(closed.groupby(["timeframe", "pattern"]).pnl_net_usdt.agg(["count", "sum", "mean"]))

Large tables are split into several Parquet files and can also be read directly with pandas, Polars or DuckDB:

import duckdb

duckdb.sql("""
    SELECT pair, timeframe, ts, open, high, low, close, volume
    FROM 'hf://datasets/Azzindani/kestrel/data/candles/*.parquet'
    WHERE pair = 'BTC/USDT' AND timeframe = '5m'
    QUALIFY row_number() OVER (PARTITION BY pair, timeframe, ts ORDER BY id) = 1
""").df()

Disclaimer

All trades are simulated. Simulated fills differ from real exchange fills. Nothing here is financial advice.

License

CC BY 4.0. Use it freely and credit the source.