Moving-Average Trend Rules Without the Mystery

Start with the arithmetic

A simple moving average gives equal weight to each observation in a chosen window and replaces the oldest observation when a new one arrives. The NIST single moving-average explanation presents this rolling calculation. A directional rule adds a comparison to that statistic; the average by itself does not instruct anyone to take a position.

For a hypothetical three-observation window containing 98, 100 and 102, the mean is 100. If 104 arrives, the next window is 100, 102 and 104, giving 102. The mean increased by two because the replacement value exceeded the departing value by six, and six divided by three is two. That identity explains every update of this equal-weight average.

Compare price with its trailing mean

One rule assigns +1 when the current level exceeds a trailing mean, -1 when it falls below, and zero at equality. Whether the current observation is included in the mean must be stated. Including it is permissible after that observation is available; using a value before its publication is a different issue. Window membership and timing are separate decisions.

With hypothetical levels 100, 103, 102, 105 and 104, the five-observation mean is 102.8. The latest level exceeds it by 1.2, so this convention gives +1. That gap has units of the reference series. It is not a percentage confidence score, an expected movement or a quantity of contracts, and it should not be displayed as one.

Compare a fast mean with a slow mean

A crossover rule replaces the latest individual level with a shorter average, then compares that fast mean with a longer slow mean. For the same five values, a two-observation fast mean is 104.5 while the five-observation slow mean is 102.8. The difference is 1.7. A positive difference produces +1 under a sign-based crossover convention, independently of the gap's magnitude.

The two averages share observations, so they are not independent votes. Recent values enter both calculations but receive different weights. A two-observation mean gives each of its values one half of the weight; a five-observation mean gives each one fifth. Calling their agreement two separate confirmations would overstate how much distinct information the rule actually contains.

Understand disagreement with endpoint rules

A lookback change uses two endpoints, while an average includes intermediate values. Consider 100, 130, 130, 130 and 110. The four-interval endpoint change is +10, yet the five-observation mean is 120 and the latest level is ten below it. The endpoint rule says positive and the price-versus-mean rule says negative without either calculation being inconsistent.

The disagreement describes the path: the latest level is above the starting point but below the intervening elevated levels. It does not identify a superior rule. A different hypothetical path with identical endpoints could produce a different average and remove the disagreement. This is why an endpoint-only summary cannot reconstruct a moving-average label without the observations inside the window.

Specify warmup and equality

A five-observation average needs five eligible values under a full-window convention. Computing an average from the first two values and labelling it a five-observation statistic changes the early rule. The pandas rolling documentation exposes minimum-observation and alignment choices explicitly. Those choices belong in the method description even when the calculation is performed without software.

Equality also needs a policy. Assigning zero, keeping the previous label, and treating equality as positive generate different sequences on flat data. For five identical values of 100, both means and the latest level equal 100. A neutral-at-equality rule produces zero. A retain-previous rule needs an existing state and an initialization convention before its answer can be determined.

Distinguish smoothing from delay

A trailing average reacts gradually to a one-time change because earlier values remain in its window. An explicit two-observation delay postpones the use of whatever comparison has already been calculated. These mechanisms are not interchangeable. Increasing the average length changes the statistic itself; adding a delay preserves the statistic but changes which row is allowed to display its resulting state.

A centered average is another distinct object: its placement may use observations later than the date receiving the average. It can be useful for retrospective description but must not be mistaken for a contemporaneously available signal. Clear labels should say trailing or centered and identify the decision cutoff. Visual smoothness does not demonstrate that the information was available at that time.

Interpret a crossover conservatively

A crossover can occur because the newest value changes, because an old value leaves a window, or because both happen together. Examining the entering and departing observations often explains a surprising label better than inventing a market narrative. Rounding can also conceal a small difference, so calculations should retain precision while the display explains its rounding convention separately.

These rules are useful teaching devices because their outputs can be checked by hand. Compare their eligible labels, persistence and state changes on identical inputs and timing. Do not convert the size of a mean gap into implied predictive accuracy. Execution, contract mapping, volatility sizing and costs require additional assumptions that a moving-average line, however familiar, does not provide.

Source record

1. NIST: Single Moving Average
2. pandas: Rolling Windows, Minimum Observations and Alignment