Filtering entries of a trend-following strategy - page 2

[Kenichiro Sakamoto]

I have spent a lot of time on exactly these three questions with trend systems, and most of what I learned was negative - i.e. things that sounded right and measurably were not. That may be more useful to you than another list of filter ideas, so here it is.

1. On filtering out the flat market.

Every filter I added to a working trend system made it worse, and the way it made it worse was instructive. Filters do not remove "bad" trades evenly - they remove trades from the left tail *and* from the right tail, and a trend system lives entirely in the right tail. When I tightened entry conditions on a gold breakout system, the win rate went up and the total profit went down, because a couple of the biggest winners had started from entries the new filter rejected.

Two concrete measurements from my own tests:
- Loosening the entry gate to get more trades: total profit over four years fell by about 78%.
- Tightening it to get fewer, "higher quality" trades: profit fell as well.
The original, unfiltered gate was the optimum, and both directions were worse. The lesson I took: if a filter improves your backtest, check *which* trades it removed. If the biggest winners are gone and the equity curve is smoother, you have not improved the system - you have converted it into a different, worse one that merely feels nicer.

So before hunting for a better flat-market filter, run this experiment: sort your trades by profit, delete the top 5% from the results, and see what is left. If the strategy is unprofitable without its top 5%, then any filter that touches those trades is dangerous, no matter how good it looks on aggregate metrics.

If you still want a regime measure, ADX or an ATR-percentile is a reasonable place to start - but treat it as something to *disable* the system in obvious chop, not as an entry refiner.

2. On pyramiding (up to 10 positions along the trend).

The honest question to answer first is whether the pyramid is adding edge or adding leverage. A quick test: run the strategy with a single position and the same total risk, and compare. In my experience the pyramid usually wins in backtests mostly because it puts more money on the table during the few big trends - which is the same thing as raising risk, and it will also deepen the drawdown when a trend fails after you have added your last unit.

If you keep it, two rules helped me: never add to a position whose stop is not already at or beyond break-even (so an added unit cannot turn a won trade into a loss), and size later units smaller, not equal. And be aware that ten simultaneous longs on correlated instruments is one position, not ten - the correlation shows up exactly when the trend breaks.

3. On drawdown compression during flat periods.

The professional answer is disappointing: you mostly cannot remove flat-period drawdown from a trend system, because that drawdown is the fee you pay for the trend trades. What you can do:

- Cap risk per trade and make it a fixed fraction of balance, computed from the actual stop distance. Then drawdown scales predictably with risk, and you can pick a level you can sit through.
- Time-out trades that go nowhere (close after N bars if the move has not developed). This does not create edge, but it stops dead capital and reduces the number of small losses that accumulate in chop.
- A daily/weekly loss limit that pauses new entries. Note this is capital protection, not an improvement to the edge - it will occasionally pause you right before the trend that would have paid for the month.

What I would avoid is "risk return" schemes that increase size after drawdown to recover faster. They convert a survivable drawdown into an account-ending one on the tail event, and in every test I ran they did not improve the risk-adjusted return, only the emotional appeal of the equity curve.

One last suggestion. When you evaluate any of these changes, do not judge by profit factor and win rate alone. Look at what happened to the largest winning trades, and check the result on data you did not optimize on, and on a second broker's feed. I have watched "improvements" that raised the profit factor on one feed and destroyed the system on another - the improvement was fitted to the data, not to the market.