Should verified snapshots disclose partial indicator windows and initialization?

[MikasaMumei]

I would like to confirm the intended output contract before proposing any change to calculations or indicator availability.

On main 1394a3f, with stockstats 0.6.9, a verified snapshot with synthetic closes 1 through 81 returns close_200_sma = 41.00. This matches the dependency's partial-window behavior. With 200 rows it is 100.50, and with 201 rows it is 101.50. The snapshot does not disclose the 81-observation coverage; its recent-closes table displays at most 30 rows.

An offline audit of all 13 Yahoo indicator names found several different initialization contracts in the installed dependency's upstream source:

Indicators	Observed behavior with short history
50/200 SMA	Means over the available observations before the fixed window is full.
Bollinger middle/upper/lower (20)	Partial-window mean and standard deviation; with one row the middle is numeric but the bands are NaN.
VWMA (14)	Partial-window price/volume sums.
10 EMA and MACD/signal/histogram (12,26,9)	Exponential smoothing starts early; the periods are not finite history requirements or convergence guarantees.
RSI/ATR (14)	Smoothed initialization; RSI's first value is explicitly set to 50.
MFI (14)	The first 14 rows are explicitly set to 0.5, including the 14-row boundary.

VWMA and MFI are in the Yahoo indicator tool catalog, but not in the default snapshot set. These are observations of stockstats 0.6.9, not assumptions about Alpha Vantage or all allowed dependency versions.

Minimal offline reproduction in an existing TradingAgents environment:

from unittest.mock import patch

import pandas as pd

from tradingagents.dataflows.vendors.yahoo import snapshot

closes = list(range(1, 14))
frame = pd.DataFrame({
    "Date": pd.bdate_range(end="2026-05-08", periods=len(closes)),
    "Open": closes,
    "High": [v + 1 for v in closes], "Low": [v - 1 for v in closes],
    "Close": closes, "Volume": 1000,
})
with patch.object(snapshot, "load_ohlcv", return_value=frame):
    print(snapshot.build_verified_market_snapshot(
        "EXAMPLE", "2026-05-08",
        indicators=("close_50_sma", "close_200_sma", "boll", "boll_ub", "boll_lb"),
    ))

The two SMAs and Bollinger middle each render as 7.00 despite their different requested windows. The bands are also numeric. The market analyst is instructed to use this snapshot for exact indicator claims, so the distinction may matter for short-history instruments. This does not establish an effect on final model decisions.

I also compared an oscillating 201-row history with its own trailing subsets. With identical final 26 observations, Bollinger and VWMA matched, while EMA, MACD, RSI, and ATR still differed. I would therefore avoid declaring those smoothed indicators fully initialized at an arbitrary period count.

I saw that #904, which included strict-window checks, was closed as superseded by the verified snapshot. I do not read that as an explicit decision on initialization disclosure.

Would you prefer:

• A short coverage note for underfilled fixed-window indicators, keeping every value unchanged, starting with SMA and Bollinger in the snapshot? For example: “close_200_sma uses 81 closing-price observations; the 200-observation window is not full.” VWMA could receive equivalent disclosure in the Yahoo tool path if useful.
• Documentation of the existing initialization contracts, including smoothed indicators and MFI's initial values, rather than additional runtime notes?

My preference is the first option for finite rolling windows, with documentation for the other families. Coverage would follow each indicator's actual input and analysis cutoff, not calendar days or the displayed recent-row count. I would keep full-window output unchanged and avoid applying one blanket minimum-history rule to every indicator. No switch to N/A or algorithm change is proposed here; Yahoo profile averages remain a separate source.

[Robert-Study]

I'd go with your first option, applied only to finite rolling windows, plus documentation for the smoothed family.

Why the two families need different treatment

• SMA, Bollinger and VWMA over N bars are finite-window statistics. With fewer than N rows, stockstats is effectively using min_periods=1, so close_200_sma = 41.0 on 81 rows is a valid mean of 81 closes but not a 200-day SMA. That is a different quantity under the same name, so the snapshot should say so. In pandas terms, the honest output is NaN until the window is full (rolling(200, min_periods=200)), or the value plus a note. Your "keep values unchanged, add a note" proposal is the lower-risk option because it does not change any numbers the analyst already sees.
• EMA, MACD, RSI and ATR are recursive (they depend on all earlier data through a decay factor). They have no hard "full window" point. Your 201-row experiment shows this: after the same last 26 bars the values still differ. The usual rule of thumb is to treat the first ~3-5x the span as warm-up, but that is a heuristic, not a guarantee. So I would document them rather than print a coverage note against a fake threshold, and avoid one blanket minimum-history rule.
• MFI/RSI seeded at fixed constants (0.5, 50) for the first rows are placeholders, not signals. Worth saying that in the docs.