Momentum · Persistent demand and reversal risk

Observation

Do assets that rose over a past period continue to earn higher returns? Distinguish time-series momentum (an asset relative to its own history) from cross-sectional momentum (rankings among assets). If all assets rise on the same date, being top-ranked does not make an absolute upward signal identical to relative strength.

Mechanism

Gradual information diffusion, staged capital allocation, and risk-budget adjustments may sustain demand. The risk bearer faces abrupt reversals, repeated trading in flat markets, and simultaneous participant exits. A mechanism provides a research direction; it does not prove every moving average works.

Hypothesis

Teaching hypothesis: after the daily close, calculate the return over the last 20 complete days. If positive, hold one unit of risk at the next trading time; if negative or data is incomplete, stay flat. Reassess after one day. Fix the long-term OOS net-return criterion before running the experiment. Twenty days is a parameter to test, not a recommended horizon.

Data

Retain daily-bar end times, actual data-availability times, the next executable quote, costs, and asset listing/delisting records. During an exchange outage, do not treat the last price as an executable quote. A cross-sectional version must reconstruct the tradable universe at every date rather than backfill today's survivors.

Signal

Translate the sign of past returns into a target position, then estimate volatility from past data to normalize risk. Keep separate columns for the signal, risk-adjusted target, and actual fills. If estimated volatility approaches zero, apply a predeclared cap or stop; an inverse must not create unbounded size.

Backtest

Use the next executable price and include costs at every position change. Compare flat, equal-risk holding, and fixed-direction baselines. Report three predeclared regimes—rising, falling, and ranging—and register every tested lookback. If many lookbacks were tried, give the final test to the frozen candidate only once.

Falsification

Does removing a few of the best trend episodes destroy explanatory power? This is sensitivity analysis; retain the original result. Delay the signal one day, raise costs, and add gap reversals. Check whether the edge still covers execution costs. A single successful period or parameter is insufficient for promotion.

Exercise and project acceptance

Write two cards, time-series and cross-sectional, changing only the signal definition while holding clocks, risk, and fees constant. Submit parameter registration, net-return distributions, turnover, and a reversal record. A reviewer should identify when the signal became visible, the source of maximum losses, and the distinction between “prices rose” and “the strategy has an edge.”

Enter the candidate into the Strategy Object, then examine shared portfolio risks.