jit_compilation_overhead.rst

JIT Compilation Overhead

HftBacktest takes advantage of Numba's capabilities, relying on Numba JIT'ed classes. As a result, importing HftBacktest requires JIT compilation, which may take a few seconds. Additionally, the strategy function needs to be JIT'ed' for performant backtesting, which also takes time to compile. Although this may not be significant when backtesting for multiple days, it can still be bothersome. To minimize this overhead, you can consider using Numba's cache feature. See the example below.

from numba import njit
# May take a few seconds
from hftbacktest import BacktestAsset, HashMapMarketDepthBacktest

# Enables caching feature
@njit(cache=True)
def algo(arguments, hbt):
    # your algo implementation.

asset = (
    BacktestAsset()
        .linear_asset(1.0)
        .data([
            'data/ethusdt_20221003.npz',
            'data/ethusdt_20221004.npz',
            'data/ethusdt_20221005.npz',
            'data/ethusdt_20221006.npz',
            'data/ethusdt_20221007.npz'
        ])
        .initial_snapshot('data/ethusdt_20221002_eod.npz')
        .no_partial_fill_exchange()
        .intp_order_latency([
            'data/latency_20221003.npz',
            'data/latency_20221004.npz',
            'data/latency_20221005.npz',
            'data/latency_20221006.npz',
            'data/latency_20221007.npz'
        ])
        .power_prob_queue_model3(3.0)
        .tick_size(0.01)
        .lot_size(0.001)
        .trading_value_fee_model(0.0002, 0.0007)
)

hbt = HashMapMarketDepthBacktest([asset])
algo(arguments, hbt)