A walk-forward check for bond-breadth timing

The first fit was discarded

The archive first combined utilities with market breadth. That combination looked strong on a fitted 10-year daily sample, then failed on held-out data and was discarded as curve-fitting.

After that failure, a continuous T-bond futures series replaced the utilities series. Because those futures began trading in 1977, earlier history could not be examined.

What the replacement rules used

The mechanical trading system takes the S&P 500, continuous T-bond futures, and daily NYSE advancing and declining issues and volume as inputs. Those counts are converted into issue and volume ratios so earlier periods remain comparable after NYSE listings more than doubled from 1978 and volume later reached about 10 times the 1978 level. Market breadth is carried as advancing issues divided by declining issues, and as the matching volume ratio.

T-bond quotes are treated as price of par, so a rising smoothed futures price is read as falling yields. Daily futures are exponentially smoothed with the two-day constant 2/(1+2). T-bond direction is the smoothed bond-futures price minus its value a chosen number of days earlier.

Buy and sell conditions are specified independently. Both require persistent T-bond direction, breadth or volume-ratio thresholds, and an S&P 500 move of a set percentage from a local extreme before a close entry or exit.

The predeclared split

Fourteen parameters were searched after the sample from January 3, 1978 through August 8, 1997 was split into a January 2, 1987 to December 29, 1995 test window and held-out windows of January 3, 1978 to February 27, 1987 and January 3, 1996 to August 8, 1997.

All 14 parameters could not be searched at once. Buy and sell groups were optimized separately in batches of up to four variables, with later passes using the previous groups' selected values.

Similar results on test and held-out segments were treated as necessary confidence, not as a guarantee of later behavior. Fitted-sample results alone were described as unusable for forecasting.

A sell rule that could not exit quickly

The selected sell-direction rule required 11 consecutive down days in smoothed T-bond direction. That left about a two-week interval in which no exit could fire. Late exits in the full sample were blamed on decline-advance issue or volume ratios not triggering, including a March 29, 1994 exit 6.1% below the February 2, 1994 peak.