Point-in-Time Backtesting

The Look-Ahead Bias Problem

Most financial datasets don't tell you WHEN a data point became available. If you use a company's 2023 annual results (filed March 2024) in a January 2024 backtest, you've cheated — that information wasn't available yet. Valuein's knowledge_at field records the exact SEC acceptance timestamp for every fact, enabling rigorous PIT backtesting.

Filtering by knowledge_at

Always filter by knowledge_at <= rebalance_date when selecting signals for a historical portfolio.

from valuein_sdk import ValueinClient, ValueinError

rebalance_date = "2022-01-01"

try:
    with ValueinClient() as client:

        # PIT-safe revenue for all companies as of 2022-01-01
        revenue_pit = client.query(f"""
          SELECT entity_id, numeric_value AS revenue, knowledge_at
          FROM fact
          WHERE standard_concept = 'TotalRevenue'
            AND fiscal_period = 'FY'
            AND knowledge_at <= TIMESTAMP '{rebalance_date}'
          QUALIFY ROW_NUMBER() OVER (
            PARTITION BY entity_id ORDER BY knowledge_at DESC
          ) = 1
        """)
        print(revenue_pit.head())

except ValueinError as e:
    print(f"Error: {e}")

Survivorship-Bias-Free Universe

Never use today's S&P 500 list for historical backtests. Use get_pit_universe to get the exact constituents at each rebalance date.

from valuein_sdk import ValueinClient, ValueinError

# MCP: get the exact S&P 500 on Jan 1, 2020
# get_pit_universe(as_of_date="2020-01-01", index="SP500")

# Python SDK:
try:
    with ValueinClient() as client:
        universe_2020 = client.query("""
          SELECT r.entity_id, r.symbol, r.name, r.sector
          FROM references r
          JOIN index_membership im ON r.security_id = im.security_id
          WHERE im.index_name = 'SP500'
            AND im.start_date <= DATE '2020-01-01'
            AND (im.end_date IS NULL OR im.end_date > DATE '2020-01-01')
        """)
        print(f"S&P 500 universe on 2020-01-01: {len(universe_2020)} companies")

except ValueinError as e:
    print(f"Error: {e}")

Full Monthly Rebalance Loop

Combine PIT universe + PIT signals to build a proper factor backtest.

import pandas as pd
from valuein_sdk import ValueinClient, ValueinError

rebalance_dates = pd.date_range("2018-01-01", "2023-12-01", freq="QS")
portfolio_returns = []

try:
    with ValueinClient() as client:
        for date in rebalance_dates:
            date_str = date.strftime("%Y-%m-%d")

            # Step 1: PIT universe on this date
            universe = client.query(f"""
              SELECT r.entity_id FROM references r
              JOIN index_membership im ON r.security_id = im.security_id
              WHERE im.index_name = 'SP500'
                AND im.start_date <= DATE '{date_str}'
                AND (im.end_date IS NULL OR im.end_date > DATE '{date_str}')
            """)

            # Step 2: Latest known revenue for universe
            ids = ','.join(repr(x) for x in universe.entity_id.tolist())
            signals = client.query(f"""
              SELECT entity_id,
                numeric_value AS revenue,
                knowledge_at
              FROM fact
              WHERE standard_concept = 'TotalRevenue'
                AND fiscal_period = 'FY'
                AND knowledge_at <= TIMESTAMP '{date_str}'
                AND entity_id IN ({ids})
              QUALIFY ROW_NUMBER() OVER (
                PARTITION BY entity_id ORDER BY knowledge_at DESC
              ) = 1
            """)

            # Step 3: Rank by revenue growth (simplified)
            signals['rank'] = signals['revenue'].rank(pct=True)
            top_quintile = signals[signals['rank'] > 0.8]
            portfolio_returns.append({'date': date, 'n': len(top_quintile)})

except ValueinError as e:
    print(f"Error: {e}")