Point-in-time fundamentals: avoiding look-ahead bias

Look-ahead bias means using information on a past date that only became known later.

Results for a March fiscal year are not public on March 31. Join them by period-end date and, for the one to two months until the May release, you select stocks using future numbers.

The error is especially common in metrics that combine prices and results, such as PER and PBR.

For market cap and valuation histories, each day uses the fiscal period with the latest period end among those disclosed on or before that day.

• Disclosure date: the earliest publication time on or after the period end recorded in the period's source information. Usually the earnings summary date.
• Unknown disclosure date: the period is used from 45 days after period end (the customary deadline for earnings summaries).
• Price dates before the disclosure date: the previous period is still used. The switch happens on the disclosure date.
• Split adjustment: prices, per-share values and share counts are all on today's share basis, so metrics do not jump on split dates.

Each metric is computed from that day's split-adjusted close and the results usable on that day.

In Toyota's weekly market cap history, the share-count source (sharesPeriod) switches on the August 7, 2026 row. It is the first week-end after the Q1 earnings summary for the fiscal year ending March 2027, published on August 4, 2026.

Before that, the full-year share count for the fiscal year ended March 2026 (13,033,384,474 shares) is used; after it, the Q1-end count (11,841,833,660 shares).

PER over the same weeks (basis=fy) uses full-year results only, so epsPeriod stays at 2026-03/FY. Use basis=ttm to switch every quarter.

Request a date range and a sampling interval (interval = daily, weekly or monthly). In a backtest, attach to each decision date the latest row on or before that date.

curl -sG "https://api.zenithcode.net/stable/valuation-history" \
  --data-urlencode "symbol=7203" \
  --data-urlencode "from=2025-01-01" \
  --data-urlencode "to=2025-12-31" \
  --data-urlencode "interval=monthly" \
  --data-urlencode "basis=ttm" \
  -H "x-api-key: YOUR_API_KEY"

import requests
import pandas as pd

BASE = "https://api.zenithcode.net/stable"
HEADERS = {"x-api-key": "YOUR_API_KEY"}

def get(path, **params):
    r = requests.get(f"{BASE}{path}", params=params, headers=HEADERS, timeout=30)
    r.raise_for_status()
    return r.json()

# Daily PER / PBR built only from results disclosed on or before each date
val = pd.DataFrame(get("/valuation-history", symbol="7203", basis="ttm",
                       **{"from": "2021-01-01", "to": "2025-12-31"}))
val["date"] = pd.to_datetime(val["date"])

# Your decision dates (e.g. month-end rebalancing)
signals = pd.DataFrame({"date": pd.date_range("2021-01-31", "2025-12-31", freq="ME")})

# merge_asof attaches the latest row on or before each date: no look-ahead
panel = pd.merge_asof(signals, val.sort_values("date"), on="date", direction="backward")
print(panel[["date", "per", "pbr", "epsPeriod", "bpsPeriod"]].tail())

ZENITH Code guarantees which period is used and from when. Account for the following yourself.

• Value versions are not stored. When an annual report or a correction replaces a value, past dates use the replaced value too. It is not exactly the number visible at the time.
• The disclosure date is the first publication. Values confirmed by the annual report are used from the summary date, so for companies whose summary and annual report differ, the gap between the two dates is affected.
• Financial endpoints return the latest values. If you join income-statement yourself, using each row from provenance.announcedAt onward is conservative (later). announcedAt is the publication time of the current source, so periods replaced by the annual report carry its date.
• Prices start in January 2010. Market cap and PER histories cannot be built before that.