Momentum vs mean reversion: test the same opportunity

Trend following vs mean reversion: different conditional bets

This page compares single-instrument continuation and reversion rules. Time-series momentum uses an instrument's own history; cross-sectional momentum ranks instruments against one another. Trend following includes several ways to measure direction, such as moving-average and breakout rules. AQR's Time Series Momentum paper makes the own-history versus relative-ranking distinction and reports horizon-dependent findings. Its futures research does not establish that a five-bar chart rule is profitable.

A matched experiment that changes only direction

The following is an educational experiment specification, not a recommended parameter set or a backtested strategy. Use one standard candlestick chart with a positive-priced instrument, one feed, and a fixed session. At completed bar t, calculate M as the arithmetic mean of the preceding 20 closes, excluding bar t. Define d = close(t) / M − 1. A bar is eligible when |d| is at least 2% and the shared experiment is flat. Equality qualifies; a smaller deviation produces no trade.

• This design isolates the sign of a position after a shared deviation signal. It does not compare every momentum strategy with every mean-reversion strategy. A momentum signal based on past returns would be a separate experiment.
• Retain the signal-time mean to explain the original hypothesis. A moving mean can catch up with a flat price: if price stays at 104 while the mean rises from 100 to 104, the gap closes without a profitable short price move.
• For a deployable strategy, define risk exits and position limits separately, then compare those full designs. Scheduled exits in this diagnostic exercise leave interim losses uncapped and are not a live risk policy.

Same signal, four possible outcomes after costs

Hypothetical prices: the prior mean is 100 and the signal close is 103, so d is +3%. Both versions enter at the next open of 104. Continuation is long; reversion is short. Each uses 10,400 of initial notional, or 100 units. For this arithmetic only, charge a fixed total round-trip cost of 0.40% of entry notional, or 41.60, to each version. This flat cost is not a claim about any broker or a TradingView commission setting.

• For the 108 exit, long net return is (108 − 104) / 104 − 0.004 = 3.446%; short net return is (104 − 108) / 104 − 0.004 = −4.246%. Returns use initial notional, not account equity. Displayed figures are rounded to two decimals.
• Before costs the matched long and short payoffs cancel. After these costs they sum to −0.80% per opportunity. Reversing an unprofitable strategy does not automatically create a profitable one.
• Exit prices do not reveal the path between entry and exit. A trade that finishes profitably could have had a large adverse excursion. Check intratrade risk and margin separately; a real position may not survive until the scheduled exit.

Regime changes and tail losses can overturn the ranking

A range that becomes a sustained trend can repeatedly punish a fade rule. A trend that turns sharply can punish continuation. These are failure mechanisms to test, not a claim that either strategy always has a particular return distribution. Daniel and Moskowitz document severe losses in their studied momentum portfolios around market rebounds; that result should not be assigned automatically to every single-chart trend strategy.

• Choose any regime filter using information available at the signal close. Record its lookback and threshold before evaluation. Labeling a period as a trend because of what happened after entry is useful descriptive analysis, but cannot act as a live entry filter.
• Split results by calendar period and a predeclared volatility or trend condition. Report the number of trades in each group. A tiny winning group is weak evidence for a regime-switching system.
• Test the switching rule as a third strategy on later, untouched data, including delays, turnover, and misclassifications. Do not splice together whichever standalone rule won each historical segment.
• Inspect the worst trade, drawdown duration, adverse excursion, and concentration of profit in a few trades. Repeated small reversion wins can be erased by one persistent move; frequent continuation losses can consume a later trend gain.
• Short positions add asymmetric constraints: borrowing may be unavailable, costs can vary, and a rising price has no fixed upper bound. An identical long/short fill model cannot establish real-world implementability.

Keep the benchmark and evaluation horizon visible

Evaluate both versions on the same chronological test segment, with the same warm-up and a shared final entry cutoff that leaves five full holding bars. Keep a zero-return cash reference for this simplified experiment and an always-long control using exactly the same entry and exit windows. Also show full-period buy-and-hold separately; its continuous exposure is different from the intermittent strategies.

• Report net return, maximum drawdown, exposure, trade count, average net trade, and worst trade. Compare long and short subsets to see whether a result mostly reflects the market direction.
• Record all tested lookbacks, thresholds, and holding periods. Choose parameters on an earlier segment, freeze them, and evaluate on a later one. The holdout becomes part of model selection if you keep revising rules after seeing it.
• Stress costs upward and check delayed fills. A small gross edge can disappear even if the signal direction is correct more often than not.
• For TradingView, verify actual order timestamps against the specification. Default strategy market orders normally fill on the next available tick, commonly the following bar open after a close calculation. Different calculation and order-processing settings change that behavior.

Turn the experiment into two inspectable TradingView runs

Ask the Pine Script AI Coding Agent to draft one strategy with a continuation/reversion selector and the shared rules above. Request a visible log of signal time, reference mean, direction, entry, and scheduled exit. Review and syntax-check the generated source, then verify timestamps and quantities in TradingView before interpreting returns. The agent drafts code; it does not establish a trading edge.

• Export each run's TradingView Strategy Tester List of trades CSV with its settings recorded. Import the files separately into Backtest Deep Report to examine trade distributions and drawdowns, then compare the same metrics. This is a manual comparison workflow, not a claim of an automatic paired experiment.
• Backtest Deep Report expects trade-list CSV data, not raw OHLC candles. Its imported trade analysis cannot reconstruct every missing intrabar equity observation. Check current upload access in the app before planning the workflow.