Skip to main content
Conduit is a point-in-time data API. The core differentiator is that every observation is bi-temporal: it records both what period a value describes and when that value became known. Keeping those two axes separate is what lets you answer questions like “what was the latest US CPI reading that I could have seen on 2026-03-15?” without contaminating the answer with data that was published later. This page explains the three time axes, how to filter on each, and the honest limits of the current as_of implementation.

​
Three axes, never conflated

Described period

period, period_start, period_end. The slice of real-world time the value is about: May 2026 CPI, Q1 2026 GDP, the trading day for a daily series.

Knowledge time

start_date, end_date. Bounds on observed_at: when the value became known to Conduit. This is the axis that prevents lookahead.

Vintage cutoff

as_of. Returns the latest vintage known on or before a timestamp. Revisions are retained, never overwritten, so you can replay history.
The mistake most data APIs make is collapsing these into one timestamp. A figure released in June that describes May, then revised in July, has at least three distinct timestamps attached to it. Conduit keeps them apart so each query means exactly one thing.

​
1. The described period

The period is the real-world interval a value measures. It does not move when the value is revised.
  • period is a convenience filter and accepts YYYY, YYYY-MM, or YYYY-Qn.
  • period_start and period_end are the explicit bounds on the response object.
# All 2026 readings for US headline inflation, by described period
curl -G "https://data.quantoraresearch.com/v1/public/observations" \
  -H "x-api-key: $CONDUIT_API_KEY" \
  --data-urlencode "country=USA" \
  --data-urlencode "indicator=cpi_inflation_yoy" \
  --data-urlencode "period=2026"
period=2026-Q1 selects observations whose described period falls in the first quarter. Use it for quarterly series such as GDP; use period=2026-05 for monthly series such as CPI.

​
2. Knowledge time (the no-lookahead axis)

observed_at is the moment a value became known to Conduit. You bound it with start_date and end_date. This is the axis you filter on when you care about what was knowable, not what period it covers. 
# Everything that became known in March 2026, regardless of which period it describes
curl -G "https://data.quantoraresearch.com/v1/observations" \
  -H "x-api-key: $CONDUIT_API_KEY" \
  --data-urlencode "country=USA" \
  --data-urlencode "indicator=cpi_inflation_yoy" \
  --data-urlencode "start_date=2026-03-01T00:00:00Z" \
  --data-urlencode "end_date=2026-03-31T23:59:59Z"
A May CPI release that lands on 2026-06-11 has period = 2026-05 but observed_at in June. Filtering on period and filtering on knowledge time answer different questions. Pick the one that matches your intent.

​
3. Vintage cutoff with as_of

as_of asks: “give me the data as it stood at this moment.” It returns, for each matching series, the latest vintage known on or before the supplied timestamp. Because Conduit retains revisions instead of overwriting them, two as_of queries with different cutoffs can return different values for the same period. 
curl -G "https://data.quantoraresearch.com/v1/observations/latest" \
  -H "x-api-key: $CONDUIT_API_KEY" \
  --data-urlencode "country=USA" \
  --data-urlencode "indicator=cpi_inflation_yoy" \
  --data-urlencode "as_of=2026-03-15T00:00:00Z"
The same query with as_of=2026-07-01T00:00:00Z may return a different number for the same described period, because a revision became known in between. Both answers are correct, each for its own cutoff.
Honest caveat about as_of. Today, as_of approximates the ingestion timestamp: the moment Conduit recorded a value. It is not yet a full provider-vintage reconstruction. If a provider published a figure before Conduit ingested it, the as_of cutoff reflects when Conduit knew, not the instant the provider released it. Full provider-vintage reconstruction is planned future work. For most backtests this distinction is small, but if your strategy is sensitive to intraday release timing, account for it.

​
Why this matters for backtests

Lookahead bias is the silent killer of historical research. If your backtest on 2026-03-15 quietly uses a CPI value that was only revised into existence in July, your results are fiction. Conduit removes that risk in two ways:

Pin the cutoff

Run each historical query with as_of set to the simulated decision time. You receive only the vintage that was known on or before that moment.

Or bound knowledge time

Alternatively, constrain end_date to the decision time. Anything that became known later is excluded by construction.
import os, requests

API = "https://data.quantoraresearch.com"
HEADERS = {"x-api-key": os.environ["CONDUIT_API_KEY"]}

def latest_known_at(decision_time: str):
    """The CPI vintage that was knowable at decision_time. No lookahead."""
    resp = requests.get(
        f"{API}/v1/observations/latest",
        headers=HEADERS,
        params={
            "country": "USA",
            "indicator": "cpi_inflation_yoy",
            "as_of": decision_time,
        },
    )
    resp.raise_for_status()
    return resp.json()["data"]

# Walk the decision dates of your backtest; each call is point-in-time clean.
for t in ["2026-01-15T00:00:00Z", "2026-02-15T00:00:00Z", "2026-03-15T00:00:00Z"]:
    print(t, latest_known_at(t))
Treat period as the x-axis of your chart and as_of (or end_date) as the lens you view it through. The period tells you which point you are plotting; the cutoff tells you which vintage of that point you are allowed to see.

​
Quick reference

AxisParametersAnswersUse when
Described periodperiod, period_start, period_endWhich real-world interval the value is aboutYou want a specific month, quarter, or year
Knowledge timestart_date, end_dateWhen the value became knownYou need a clean as-of-then view or a release window
Vintage cutoffas_ofThe latest vintage known on or before a timestampYou are reconstructing a point-in-time snapshot for a backtest
All three axes compose. You can request a specific period while pinning as_of to see how that exact reading looked at an earlier moment. Unknown query parameters fail closed with bad_request, so a typo in a time parameter is rejected rather than silently ignored.

Provenance

Trace any observation back to its named source and ingestion run.

Freshness and liveness

Understand the freshness label and how silently-frozen feeds are detected.