DataFeed

DataFeed: the only object that hands out data during a run.

Everything a model sees while being evaluated flows through here, bound to the simulation clock (asof). Availability rules live on the Field; the feed enforces them:

• actuals — values stamped t are served iff t + availability_lag <= asof;

• forecasts — runs issued at i are served iff i + availability_lag <= asof, and per valid_time the latest such run wins;

• statics — always served.

With the feed as the sole gate, look-ahead leakage is impossible by construction rather than by runner discipline.

class emflow.data.feed.DataFeed(dataset: Dataset)

Bases: object

Point-in-time access to a Dataset.

history(asof, name: str, window=None, strict=False) → DataFrame

Actual values of field name knowable at asof.

For bitemporal fields (knowledge_col set) the filter runs on the knowledge axis and, per measurement timestamp, the latest revision already knowable wins — a backtest sees preliminary values exactly as a live participant did. The knowledge column is dropped from the returned frame.

window (optional, anything pd.Timedelta accepts) limits the result to the trailing window before the availability cutoff — an efficiency knob, never a leakage one.

strict excludes the cutoff instant itself (< instead of <=). Training views use this so a training boundary placed exactly on the first scored target never serves it.

forecasts(asof, name: str, columns=None) → DataFrame

Latest forecast run of field name knowable at asof.

Returns a frame indexed by valid_time where, for each valid_time, the value comes from the most recent run with issue_time + availability_lag <= asof. An issue_time column records which run each row came from.

static(name: str) → DataFrame

actuals_between(name: str, start, end, asof) → DataFrame

Actuals of name in [start, end] — only if knowable at asof.

Used by environments to settle scores: returns the slice restricted to what the availability rule allows at asof (so an env can only score an origin once its targets have actually arrived).

view(asof, strict=False) → TimeView

A feed bound to a fixed clock time (what observations wrap).

strict=True makes the boundary exclusive — used for training views.

class emflow.data.feed.TimeView(feed: DataFeed, asof, strict=False)

Bases: object

A DataFeed frozen at one clock time.

This is the whole world a model gets to see: observations and training views are TimeView s. There is deliberately no way to reach the underlying dataset or move the clock from here.

history(name: str, window=None) → DataFrame

forecasts(name: str, columns=None) → DataFrame

static(name: str) → DataFrame

property fields