Your Python backtests are slow. Your trading system is slower. I make both fast.

# Nested loops per asset per timestep
for t in range(signal_sigma_window_size, n_obs - 1):
    sigmas = sig.iloc[t - signal_sigma_window_size:t].std()

    for j in range(n_assets):
        # Recompute same operations repeatedly
        sigma_j = sigmas.iloc[j]
        raw_signal = signal.iloc[t, j] / sigma_j

        # Apply signal filters element by element
        if raw_signal > entry_threshold:
            position[t+1, j] = 1.0
        elif raw_signal < -entry_threshold:
            position[t+1, j] = -1.0
        else:
            position[t+1, j] = 0.0

# Runtime: 35.24s for 3000 timesteps × 100 assets

# Use Polars rolling operations (Rust-based)
sigma_exprs = [
    pl.col(col)
    .rolling_std(window_size=signal_sigma_window_size)
        .shift(1) for col in signal_pl.columns
]
sigma = signal_pl.select(sigma_exprs).to_numpy()

# NumPy vectorized operations (entire matrix at once)
normed_signal = signal.to_numpy() / sigma

position_path = np.where(
    normed_signal > entry_threshold, 1.0,
    np.where(normed_signal < -entry_threshold, -1.0, 0.0)
)

# Runtime: 0.057s → Identical API, 615× speedup