You build a trend-following system. You test it on 20-day momentum and get solid results. You try 200-day momentum and it still works. Then you shorten the lookback to five minutes and everything falls apart. You extend it to three years and the edge disappears again. The trending window is the range of time horizons where price momentum actually persists. Outside that range, prices revert. Most traders pick their lookback periods by convention or by copying someone else’s settings. A 2025 study gives an empirical reason to stop guessing.
What the Research Says About the Trending Window
Safari and Schmidhuber published “Trends and Reversion in Financial Markets on Time Scales from Minutes to Decades” in Physica A: Statistical Mechanics and its Applications in 2025. They studied autocorrelation structures across equities, bonds, commodities, and currencies at horizons from minutes to decades. Their central finding: prices tend to trend at horizons from roughly one hour to about one year. Below that range, microstructure noise and bid-ask bounce create mean-reverting patterns. Above it, prices tend to revert toward long-run equilibrium values.
This is not a trading system. It is a statistical property of financial returns. The autocorrelation of returns is positive within the trending window and turns negative outside it. The implication for anyone building trend-following strategies is direct: your lookback period should fall inside this window, or you are likely trading against the grain of the data.
Why Short Lookbacks Break Down
At very short time horizons (seconds to a few minutes), price behavior is dominated by market microstructure. Bid-ask bounce alone creates negative autocorrelation in tick-level returns. A price that just moved up because it crossed from the bid to the ask is statistically more likely to drop back to the bid on the next tick. That is not a trend reversing. It is the plumbing of the market creating noise.
I run most of my scans on daily and weekly bars. When I have tested the same momentum signals on one-minute bars, the results degrade sharply. The signal-to-noise ratio collapses. You can sometimes recover an edge at very short horizons with execution-speed advantages, but that is a different game than trend following.
Where most traders get this wrong: they shorten lookbacks to “catch moves earlier” without realizing they are moving out of the trending regime entirely. A 3-bar momentum signal on a 1-minute chart is not a faster version of a 3-bar momentum signal on a daily chart. They measure fundamentally different things.
Why Very Long Lookbacks Also Fail
At the other extreme, horizons beyond about one year start showing mean reversion. This makes intuitive sense. A stock that has tripled over five years has, in many cases, already priced in the growth story. Valuations stretch. Competition arrives. The further you extend your lookback, the more you are measuring a move that has already happened rather than one that is still in progress.
The academic momentum literature has documented this separately. Jegadeesh and Titman’s classic work showed cross-sectional momentum strongest at 3-to-12-month horizons, with long-term reversal kicking in after about a year. The Safari and Schmidhuber paper extends this to time-series momentum and confirms the pattern holds across asset classes, not just equities.
I used to run a 252-day (one year) rate of change filter and noticed that extending it to 504 days did not improve screening results. It made them worse. The paper explains why: at two-year horizons, you are already in the reversion zone.
The Trending Window Across Asset Classes
One of the more useful findings is that the trending window is not unique to stocks. Safari and Schmidhuber show similar autocorrelation patterns in bonds, commodities, and FX. The boundaries shift slightly by asset class, but the general shape holds: positive autocorrelation in the intermediate range, negative at the extremes.
This matters for anyone doing multi-timeframe analysis. If you align your timeframes within the trending window, your signals reinforce each other. If your higher timeframe sits in the reversion zone (say, a five-year trend filter), it may actively contradict the signal from your intermediate timeframe.
Where this goes wrong in practice: traders use a weekly chart for “the big picture” and a daily chart for entries, which is fine. Both sit inside the trending window. But adding a monthly chart with a 10-year lookback for “confirmation” can introduce noise from the reversion regime.
Practical Lookback Settings for Trend Signals
The research does not prescribe exact lookback periods. It identifies a regime. Within that regime, specific settings still depend on your holding period, the asset you trade, and your tolerance for lag. But you can use the trending window as a boundary check.
For swing traders working daily charts, the sweet spot falls between roughly 10 and 200 trading days. A momentum indicator with a 14-day lookback sits well inside the trending window. A 200-day moving average filter is near the upper boundary but still within the zone.
For intraday traders on hourly charts, lookbacks down to about one hour (or the equivalent number of bars on your chart) remain inside the trending regime. Much shorter than that, and you cross into microstructure noise.
The one mistake I see most often: traders optimize a lookback period without checking whether the optimized value sits inside the trending window. You can curve-fit a 3-minute momentum signal to look profitable on historical data. But you are fitting to noise, not to a persistent statistical property.
How This Connects to Mean Reversion Strategies
The flip side of the trending window is equally important. If prices revert at very short and very long horizons, mean reversion strategies should target those zones. A detrended price oscillator looking for reversion works differently depending on the timeframe you apply it to.
At tick-level and sub-minute horizons, mean reversion is partly an artifact of microstructure. At multi-year horizons, it reflects genuine economic forces: valuation normalization, competitive dynamics, and capital reallocation. The mechanism is different, but the statistical signature is the same: negative autocorrelation.
This does not mean every short-term or long-term trade is a mean reversion trade. It means the default statistical tendency at those horizons favors reversion. Trend signals applied outside the window are swimming upstream.
Setting Lookback Boundaries in Your System
Here is how I apply this in practice. Before adding any trend filter or momentum signal to a screen, I check two things. First, does the lookback period fall between roughly one hour and one year of trading time? Second, does the signal improve out-of-sample when I constrain it to that range?
If a momentum signal tests best at a 5-day lookback on daily bars, that is roughly one trading week. Well inside the window. If it tests best at a 600-day lookback, I treat the result with suspicion. It may be capturing a specific historical episode rather than an ongoing statistical tendency.
The research also suggests why time-series momentum strategies typically use 1-to-12-month lookbacks. That range was not chosen arbitrarily. It sits squarely inside the empirically identified trending window.
For backtesting, the trending window provides a useful sanity check. If your system requires a lookback period outside the 1-hour to 1-year range to show an edge, you should ask whether you are exploiting a real pattern or an artifact.
Limitations Worth Knowing
The trending window is a statistical average across markets and periods. It does not guarantee that every asset trends at every horizon within the window. Individual stocks can mean-revert at monthly horizons after earnings disappointments. Commodities can trend for years during supply dislocations.
The boundaries are also approximate. “Roughly one hour” and “roughly one year” are not sharp cutoffs. The autocorrelation does not flip from positive to negative at a precise bar count. It transitions gradually. Treat the window as a zone, not a line.
Finally, the window describes unconditional behavior. During crisis periods, correlations compress and the usual autocorrelation structure can break down temporarily. A trend signal that works well in normal markets may behave differently during a liquidity crisis regardless of its lookback period.
Why Convention Got the Lookback Roughly Right
The most popular moving averages in trading (20-day, 50-day, 200-day) all fall inside the trending window. The standard momentum lookbacks (10-day, 14-day, 21-day) do too. The classic cross-sectional momentum strategy uses 12-month returns with a 1-month skip, landing at the upper edge of the window.
This is not a coincidence. Traders have been optimizing lookback periods through trial and error for decades. The settings that survived are the ones that happened to sit inside the zone where trending actually persists. The Safari and Schmidhuber paper gives a statistical explanation for why those specific settings work and why more exotic choices tend not to.
The practical takeaway: you do not need to abandon your current lookback settings. You do need to stop extending them far outside the trending window in search of a “better” signal. The evidence says the signal is not there.
Educational content only. Not investment advice. Trading involves risk. You are responsible for your decisions.