Lookback, Decision Frequency and Holding Period

Keep four clocks separate

A lookback selects the history used by a formula. A decision schedule determines when the formula is evaluated. An implementation lag separates the computed value from its permitted use. A holding convention determines when an existing state can change. Combining these choices into a phrase such as monthly trend leaves substantial ambiguity, because monthly can describe any one of them.

For example, a rule could examine 20 observations, be recomputed after every new observation, delay each result by two observations and retain its state until the next delayed signal changes. Nothing in this specification imposes a 20-observation holding period. It describes a rolling information window, not a promise about the duration of a future state.

Count intervals and observations correctly

A four-interval price change compares the current level with the level four positions earlier. It therefore needs five levels. A four-observation mean needs four levels. This difference is easy to overlook when both are labelled a four-period rule. In a comparison, each formula needs its own explicit window definition and earliest eligible observation rather than a shared informal name.

Take the hypothetical sequence 100, 102, 101, 104, 103, 105. At observation five, the four-interval change is 103 minus 100, or +3. The trailing four-observation mean is the average of 102, 101, 104 and 103, or 102.5. These calculations use related but different information sets, even though their parameter labels both contain the number four.

Define what a period means

Five observations need not span five calendar days. A holiday, missing record or irregular release can make the elapsed time longer. Conversely, intraday observations can place many records inside one day. The pandas rolling documentation distinguishes fixed observation-count windows from time-based windows. This is a useful implementation distinction, not a recommendation to use a particular library.

Suppose valid observations are recorded on Monday, Tuesday, Thursday and Friday. A two-observation delay maps Tuesday's computed label to Friday's row. A two-calendar-day delay would point to Thursday instead. Neither label is meaningful until the chosen clock is declared. Filling Wednesday with Tuesday's value merely to obtain a neat calendar changes the input and can alter the calculation.

Separate scheduled decisions from state duration

A rule evaluated at every observation may remain unchanged for a long time. Its observed state duration is an output of the formula and path. A rule evaluated only every fifth observation is different: it deliberately ignores intermediate decision opportunities. It can therefore retain an old state after the same formula, evaluated more frequently, would already have changed its sign.

Consider hypothetical computed labels +1, +1, -1, -1, +1 across five decision opportunities. Updating at each opportunity records two sign reversals. Updating only at the first and fifth records no reversal between those endpoints. This does not show that the slower schedule discovered stability. It shows that its sampling schedule did not represent the intermediate states.

Recognize overlapping windows

Adjacent 20-observation windows share 19 observations, or 95% of their inputs. A sequence of 100 calculated values is therefore not equivalent to 100 independent experiments. The overlap is not a calculation error; it is an intended consequence of rolling analysis. It becomes misleading only when the number of rows is described as independent evidence about the rule.

An overlapping holding convention creates another layer. Imagine three equally weighted cohorts, opened on successive observations and each retained for three observations. If their directional states are +1, +1 and -1, their average state is one third. This blended value is neither the newest signal nor a distinct market observation. It follows from the explicitly chosen cohort bookkeeping.

Respect warmup and historical availability

A formula requiring 20 complete observations cannot produce a fully specified value on the tenth. Reporting an unavailable value is different from reporting zero. With a two-observation delay, the first complete 20-observation mean computed at row 20 appears as a delayed value at row 22. A 20-interval change first computed at row 21 appears at row 23.

For genuine forecasting work, Forecasting: Principles and Practice describes evaluation using only observations preceding the test observation. The same chronology is useful when explaining deterministic rules: a future record cannot repair an earlier information gap. Requiring complete windows and documenting publication times make the explanation reproducible without pretending that a descriptive exercise validates a trading strategy.

Compare horizons without selecting a winner

A short window gives recent observations more influence because older information leaves sooner. A long window retains earlier conditions longer. Which produces more state changes depends on the actual path, ties and formula; it is not a universal ordering. A long window can still change abruptly when an unusual old observation exits, even if the newest observation is uneventful.

A useful educational comparison holds the input, delay and eligible dates constant while varying one clock. Report how the resulting labels differ and disclose any unequal warmup. Avoid treating a visually attractive window as established merely because it describes one example neatly. The goal is to understand what each time choice changes, not to infer an optimal holding period from a chart.

Source record

1. pandas: Rolling Window Parameters
2. Hyndman and Athanasopoulos: Time Series Cross-Validation