Skip to content

Performance Metrics

TorchTrade provides a comprehensive set of trading performance metrics to evaluate agent performance. These metrics help assess risk-adjusted returns, drawdown characteristics, and trading efficiency.

Overview

The torchtrade.metrics module provides standard trading metrics including:

Metric Description Typical Good Value
Total Return Overall portfolio return > 0%
Sharpe Ratio Risk-adjusted return (volatility) > 1.0 (> 2.0 excellent)
Sortino Ratio Risk-adjusted return (downside volatility) > 1.5
Calmar Ratio Return per unit of maximum drawdown > 1.0
Max Drawdown Largest peak-to-trough decline < 20%
Max DD Duration Length of maximum drawdown period Shorter is better
Win Rate Percentage of profitable periods > 50%
Profit Factor Total profits / total losses > 1.5
Number of Trades Total trades executed Depends on strategy

Available Metrics

Individual Metrics

from torchtrade.metrics import (
    compute_max_drawdown,
    compute_sharpe_ratio,
    compute_sortino_ratio,
    compute_calmar_ratio,
    compute_win_rate,
)
import torch

portfolio_values = torch.tensor([1000, 1100, 1050, 900, 950, 1200])
returns = torch.tensor([0.01, -0.02, 0.03, -0.01, 0.02, 0.04])

# Max Drawdown - largest peak-to-trough decline
dd = compute_max_drawdown(portfolio_values)
# Returns: max_drawdown, max_drawdown_duration, current_drawdown, peak_value, trough_value

# Sharpe Ratio - risk-adjusted return (excess return / volatility)
sharpe = compute_sharpe_ratio(returns, periods_per_year=525600)

# Sortino Ratio - like Sharpe but only penalizes downside volatility
sortino = compute_sortino_ratio(returns, periods_per_year=525600)

# Calmar Ratio - annualized return / max drawdown
calmar = compute_calmar_ratio(portfolio_values, periods_per_year=525600)

# Win Rate - percentage of steps where reward > 0
win_metrics = compute_win_rate(returns)
# Returns: win_rate, avg_win, avg_loss, profit_factor

Usage with HistoryTracker

Every TorchTrade environment records episode data in a HistoryTracker (see State Management). After a rollout, extract the history and feed it to the metrics:

import torch
from torchtrade.metrics import compute_all_metrics

# Run evaluation rollout
obs = env.reset()
done = False
while not done:
    action = policy(obs)
    obs = env.step(action)
    done = obs["done"].item()

# Get history from environment
history = env.history  # HistoryTracker instance

# Compute all metrics
metrics = compute_all_metrics(
    portfolio_values=torch.tensor(history.portfolio_values),
    rewards=torch.tensor(history.rewards),
    action_history=history.actions,
    periods_per_year=525600,  # See periods table below
)
# Returns dict with: total_return, sharpe_ratio, sortino_ratio, calmar_ratio,
# max_drawdown, max_dd_duration, num_trades, win_rate, avg_win, avg_loss, profit_factor

Logging to Weights & Biases

import wandb

wandb.init(project="torchtrade-training")

for epoch in range(num_epochs):
    # ... training code ...

    # Evaluate and compute metrics from history (as above)
    history = eval_env.history
    metrics = compute_all_metrics(
        portfolio_values=torch.tensor(history.portfolio_values),
        rewards=torch.tensor(history.rewards),
        action_history=history.actions,
        periods_per_year=525600,
    )

    wandb.log({f"eval/{k}": v for k, v in metrics.items()})

Custom Metrics

Create custom metrics following the same pattern — accept torch.Tensor rewards or portfolio values, return a scalar:

def compute_max_consecutive_wins(rewards: torch.Tensor) -> int:
    """Maximum number of consecutive winning periods."""
    wins = rewards > 0
    max_streak = current = 0
    for is_win in wins:
        current = current + 1 if is_win else 0
        max_streak = max(max_streak, current)
    return max_streak

Periods Per Year Configuration

Choose periods_per_year based on your execute_on frequency:

Execute On Periods Per Year Calculation
1 minute 525,600 60 × 24 × 365
5 minutes 105,120 12 × 24 × 365
15 minutes 35,040 4 × 24 × 365
1 hour 8,760 24 × 365
4 hours 2,190 6 × 365
1 day 365 365

Match periods_per_year to your execute_on config setting.

Best Practices

  • Compare to benchmarks: Always compute buy & hold return as a baseline
  • Use multiple metrics: Don't rely on a single metric — check returns (Sharpe, Sortino), risk (max drawdown), and efficiency (win rate, profit factor)
  • Include costs: Set realistic transaction_fee and slippage in your environment config
  • Watch for extreme Sharpe: If Sharpe > 10 or < -10, check that periods_per_year matches your execution frequency

Next Steps

References