Ml4t Transaction Costs

Transaction Cost Feasibility

Before you wire costs into a backtest engine, ask a simpler question: does the signal have enough gross edge to pay for trading at all? If not, the strategy should die early.

The Problem

The full cost of a trade has three layers: explicit costs (commissions, fees), implicit costs (half the bid-ask spread paid on every execution), and market impact (your own order moving the price against you). Most bad strategies fail this economic screen before any engine-specific implementation details matter.

A strategy with 50 bps gross alpha and 30 bps round-trip costs has a safety margin of only 1.7x - too thin to survive estimation error.

The Pattern

WRONG

import numpy as np

# Flat cost assumption - ignores volume dependence
weights = compute_target_weights(signals)
turnover = np.abs(weights - prev_weights).sum()
costs = turnover * 0.001  # 10 bps flat - wrong for large trades
net_return = gross_return - costs

CORRECT

import numpy as np

weights = compute_target_weights(signals)
turnover = np.abs(weights - prev_weights)

# Volume-dependent square-root impact model
adv = volume_20d_mean                       # 20-day average daily volume ($)
participation = (turnover * portfolio_aum) / adv
spread_cost = half_spread                   # ~2-5 bps for liquid equities
impact_cost = 0.1 * np.sqrt(participation)  # Almgren-Chriss square-root model
total_cost = spread_cost + impact_cost      # per-asset, per-rebalance

net_return = gross_return - (turnover * total_cost).sum()

Cost Stack

Safety Margin Rule

gross_alpha_bps = 50
round_trip_cost_bps = 20
safety_margin = gross_alpha_bps / round_trip_cost_bps  # 2.5x

# Target: safety_margin >= 2.5x
# Below 2.0x: strategy is fragile to cost estimation error
# Below 1.5x: likely unprofitable in practice

Capacity Estimation

import numpy as np

# Maximum AUM before impact erodes alpha
universe_adv = adv_per_asset.sum()      # total $ ADV across universe
turnover_rate = 0.20                     # 20% monthly turnover
max_participation = 0.05                 # trade < 5% of ADV

capacity = max_participation * universe_adv / turnover_rate
print(f"Estimated capacity: ${capacity/1e6:.0f}M")

Guardrails

• Never backtest without at least spread costs - it is the irreducible minimum.
• Flat bps assumptions are only valid for very small portfolios trading liquid names.
• Higher turnover amplifies cost sensitivity - under the square-root model, 2x turnover ≈ 2.8x impact cost.
• Validate cost model against Transaction Cost Analysis (TCA) data when available.
• Crypto and options have much wider spreads than equities - use asset-class-specific estimates.

Hand-Off

If the strategy clears this feasibility screen, encode the actual commission, slippage, and impact assumptions with ml4t-cost-model. This skill is the economic go/no-go filter; ml4t-cost-model is the engine-configuration step.

Checklist

• All three cost layers modeled (commission, spread, impact)
• Market impact scales with trade size relative to ADV (not flat bps)
• Safety margin >= 2.5x documented (gross alpha / costs)
• Strategy capacity estimated with participation rate constraint
• Sensitivity analysis: results reported at 1x, 2x, and 3x base cost assumptions