Walk-forward evaluation of a noise-filtered bond-futures breakout

This archive article reconstructs a historical evaluation of a long/short breakout procedure on bond futures. The procedure entered when price crossed a lookback high or low by more than a penetration amount. Separate noise filters governed buys and sells, and a one-way ratchet held the unused extreme still while a position was open.

The evaluation did not treat a fitted in-sample curve as enough. It required unused intervals long enough to show multiple price variations, then asked whether open-trade drawdowns stayed in a similar band or widened after price dynamics changed.

Separate filters for buys and sells

Entries used a lookback high plus a penetration amount. Exits and shorts used a lookback low minus a penetration amount. While short, the buy-side highest high was not allowed to rise as the lookback window rolled. While long, the sell-side lowest low was not allowed to fall.

The unused side was held in one direction so an open position did not inherit a drifting opposite extreme as the lookback window moved.

Unused intervals instead of a fitted curve

The historical workflow treated in-sample curve-fitting as insufficient evidence. A rule set was treated as more than a concept only after out-of-sample observation long enough to show multiple price variations.

How the window rolled

Walk-forward testing used a four-year optimization window starting in 1989. That window advanced every six months so newer bond-price dynamics could enter the parameter set faster than a yearly roll. Each step optimized the rules on the in-sample window, then applied those parameters to the next unused interval.

When drawdowns left the familiar band

Outside one adverse year, later unused intervals produced open-trade drawdowns that stayed in a similar band to the original four-year test window. A change in price dynamics left the prior window's best parameters invalid and produced larger drawdowns. The historical workflow treated that widening as the practical clue that the rule set needed refresh.

What still stood in the way of live use

The historical workflow concluded that more walk-forward work on additional whipsaw years was still needed before treating the procedure as ready for live use.