Specify the input series, exact rule formulas, warmup, ties, missing-data policy and two-observation lag. Compare states and consistently defined turnover on a common eligible sample. Reproducible directional diagnostics do not constitute an executable strategy, backtest or investment performance record.
Write the question before choosing the answer
A bounded method comparison can ask how often deterministic rules disagree and how much their requested direction changes. It need not calculate investment returns. This distinction keeps the evidence aligned with the question: a state sequence describes a rule's output on stated inputs, while a trading record requires contract identity, execution, sizing, costs and account conventions that a reference-level comparison does not supply.
Freeze the rules and their parameters before examining the comparison. Record whether the exercise is descriptive or intended to evaluate a predeclared hypothesis. Trying many alternatives and reporting only the neatest chart obscures how the conclusion was selected. A transparent comparison includes the full specified set, the common sample boundaries and the reasons any observation is unavailable for one or more methods.
Give the input a durable identity
Record the source, series identifier, units, retrieval date, date range and exact input version. For example, EIA identifies RCLC1 as Cushing, Oklahoma crude oil futures Contract 1 in dollars per barrel. Its definitions describe Contract 1 as the earliest delivery contract. That is historical front-contract reference context, not the history of one unchanged delivery month and not evidence of a participant's executable fills.
A reproducibility note can link the official source and describe permitted transformations without creating a public mirror of raw history. Preserve provenance separately from publication choices and check applicable data-use terms before redistribution. No EIA observations are used in the invented numerical example below. Its small reference array exists solely to make formulas and timing hand-checkable, without researching or training an instrument model.
Define three rules precisely
Use the invented levels 100, 102, 104, 103, 101, 99, 100 and 102 at observations zero through seven. Rule A takes the sign of the current level minus the level three observations earlier. Rule B takes the sign of current level minus the trailing three-observation arithmetic mean, including the current observation. Rule C compares the trailing two-observation mean with the trailing four-observation mean, both including the current observation.
For all three rules, positive differences map to plus one, negative differences to minus one and exact ties to zero. Require complete windows. At observation three, A compares 103 with 100 and is positive. B compares 103 with the mean of 102, 104 and 103, which is exactly 103, so it is neutral. C compares 103.5 with 102.25 and is positive.
Delay values by observations, not calendar labels
Apply a two-observation lag after computing each raw state. The states calculated at observation three therefore become eligible at observation five: plus one, zero and plus one. At observation four, all three raw states are negative, becoming eligible at observation six. At observation five, they are again all negative, becoming eligible at observation seven. The shared delayed sample in this example contains three observations.
Pandas documents both period-based shifting and date-index shifting with a frequency argument. Those operations should not be confused: moving calendar labels is not the same specification as making an earlier value available two eligible rows later. Write the mapping explicitly as eligible state at t equals raw state at t minus two. Missing rows and calendar closures require a separately declared eligibility policy.
| Eligible observation | A: lookback change | B: level versus mean | C: fast versus slow |
|---|---|---|---|
| 5 | +1 | 0 | +1 |
| 6 | -1 | -1 | -1 |
| 7 | -1 | -1 | -1 |
Compare only like-for-like transitions
On the shared delayed sample, sum absolute changes between adjacent eligible states, excluding any assumed initial acquisition or final liquidation. Rules A and C each contribute two units when they reverse from plus one to minus one, then zero. Rule B contributes one unit when it moves from zero to minus one, then zero. The respective turnover proxies are therefore two, one and two.
That result says only how these rules change state on this invented array under this boundary convention. It does not establish costs, superiority or investment outcomes. Including an initial position or using a different window could change the totals. Keep the observation count and transition count visible: three eligible observations contain two adjacent transitions, not three independent opportunities to change from an observed prior state.
Make window and missing-data choices explicit
Pandas' rolling-window documentation distinguishes observation-count windows from elapsed-time windows and provides an explicit minimum-observation setting. Those choices belong in the method definition. A partially populated mean is not equivalent to a complete trailing window merely because both produce a number. Centred or forward-looking windows also require special care because a historical label can conceal the inclusion of observations that arrived later.
Do not backfill a missing price from a future observation or convert an unavailable signal into a neutral state without declaring that transformation. Validate chronological order and duplicate dates before calculating. A basic hand check should include a constant array, a rising array, an exact tie and insufficient history. These are arithmetic correctness checks, not independent samples proving that a method predicts market direction.
Publish an evidence-sized conclusion
A complete comparison record includes frozen inputs, formulas, parameters, timing, valid counts, state outputs and the turnover definition. It should distinguish data errors from genuine methodological differences. When two implementations disagree, inspect the first differing observation and its complete input window. Aggregate agreement percentages can hide a single boundary mistake that propagates through every later output or changes the interpretation of the sample.
The conclusion should remain limited to what was calculated: deterministic directions and their changes under a documented information policy. A separate hypothetical risk or cost budget can explore additional assumptions, but cannot be presented as observed investment performance. Reproducibility means another reader can obtain the same defined result. It does not, by itself, make that result predictive, investable or representative of any live production process.
Research checkpoints
- Freeze formulas, ties, complete-window requirements and the two-observation lag.
- Compare a shared eligible sample with an explicit turnover boundary convention.
- Reference-series reproducibility does not prove executable fills or investment performance.
Source record
Related methods
Lookback, Decision Frequency and Holding Period
Separate the history used to form a signal from its update schedule, delay and eventual duration.
Moving-Average Trend Rules Without the Mystery
Work through price-versus-mean and fast-versus-slow rules, including ties, warmup and sensitivity to window membership.
Signal Timing and Lookahead: Follow the Information Clock
Trace observation, publication, calculation and delayed use, with practical examples of leakage and missing-data traps.
Turnover and One-Way Costs: Count Every Change Once
Define a directional turnover proxy, handle reversals and roll legs, and translate explicit one-way cost assumptions into a budget.
Published 16 October 2026. Illustrations and reference diagnostics do not establish strategy profitability or execution quality.