Reversals and Whipsaw: What Changing Signals Actually Show

Define the event before counting it

A market reversal describes a change in the path of prices. A signal reversal describes a change in a rule's classification. The two need not happen at the same observation. Whipsaw usually refers to repeated state changes that fail to develop into a sustained direction, but the term needs a precise operational definition before it can support a comparison.

One descriptive definition counts transitions directly between +1 and -1. Another includes moves into and out of zero. A third requires a return to the previous sign within a specified number of observations. Each measures something different. Choose the definition before inspecting examples, and avoid retrospectively calling only inconvenient changes whipsaws while excluding similar changes that look attractive later.

Work through a boundary crossing

Use a deliberately fixed hypothetical reference boundary of 100 and levels of 101, 99, 101, 99 and 101. A rule comparing each level with that boundary produces +1, -1, +1, -1 and +1. There are four sign reversals across four transitions. The first and final levels are identical, demonstrating how endpoints can conceal considerable classification activity.

Define a directional turnover proxy as the sum of absolute changes in the labels. Each +1 to -1 transition contributes two units, so this sequence contributes eight units after the initial state. Including a starting move from zero to +1 adds one, giving nine. Neither convention is wrong, but the inclusion of initialization must be disclosed for comparisons.

Smoothing changes the response

A trailing mean mixes new observations with older ones. The NIST moving-average discussion explains the rolling arithmetic underlying this smoothing. For hypothetical levels 99, 101 and 99, the three-observation mean is about 99.667. Replacing the first 99 with 101 raises the mean to about 100.333, less than the two-point difference between the replacement values.

That damped movement does not eliminate all sign changes. Comparing the latest value with a moving boundary can still alternate frequently, and an old extreme value leaving the window can move the boundary abruptly. The relevant question is how the complete rule responds to a defined path, not whether its average line looks smoother than the underlying observations.

Examine thresholds as explicit alternatives

A hypothetical hysteresis rule can retain its prior state until the reference level exceeds 102 or falls below 98, with a declared initial state of zero. The earlier 101 and 99 sequence triggers neither boundary, so the state stays zero. This example demonstrates the rule's arithmetic only. The thresholds are invented for explanation and are not calibrated recommendations.

Now consider levels 100, 101, 102 and 103. With strict greater-than comparison, the positive state first appears at 103, while a simple above-100 rule became positive at 101. The wider boundary avoided the first example's oscillation but delayed the second example's response. This trade-off follows from the definitions; it does not require a historical performance claim.

Keep momentum crash evidence in context

Daniel and Moskowitz's academic paper Momentum Crashes studies a distinct phenomenon involving momentum portfolios and sharp market rebounds. Its analysis is a useful warning against assuming momentum has uniform behavior. It should not be relabelled as evidence that every simple futures trend rule has the same exposure, timing or vulnerability as the portfolios in that study.

A rapid reversal can interact with an old directional state, a delayed update and a sizing rule simultaneously. A descriptive signal table isolates only part of that interaction. Explaining which component is under discussion is more informative than attaching a dramatic historical label to every alternating sequence. The fictional examples here establish mechanical possibilities, not frequencies or expected outcomes.

Translate activity into costs only with assumptions

A label-change proxy has no currency unit. To construct a hypothetical cost illustration, suppose each unit represents a $10,000 change in notional exposure and each one-way transaction costs two basis points of that amount. Eight units then imply $80,000 of transacted notional and $16 of assumed costs. This is arithmetic under stipulated inputs, not an observed trading bill.

Real contract turnover can also arise without a directional reversal, such as a roll or a size adjustment. Conversely, a label reversal need not map to a fixed number of contracts. A transparent cost budget distinguishes those components and avoids treating directional changes as a complete execution ledger. Spread conditions, order size and actual fills remain unspecified in this illustration.

Compare paths without hindsight repair

Inspect several hypothetical shapes: a steady climb, alternating observations, a single step followed by stability, and a rise followed by a sharp reversal. Apply the same window, tie rule and delay to each. This reveals what the formula is designed to represent. It does not select parameters or claim that these simple shapes occur with known probabilities in real markets.

When reporting state changes, identify eligible observations and any initialization exclusion. Also show persistence or neutral time so a low reversal count cannot hide a rule that rarely leaves zero. The responsible conclusion is about responsiveness and activity under stated assumptions. Fewer reversals alone neither proves better risk control nor establishes that a strategy would perform better after costs.

Source record

1. NIST: Single Moving Average
2. Daniel and Moskowitz: Momentum Crashes, Columbia Business School