Walk-forward evaluation of gold-index bond-fund rules

One close-to-close long-or-short procedure

The design is a fully specified long-or-short procedure. It uses a daily gold-and-silver mining-stock index as the leading input and a long-term government bond fund as the traded instrument. That pairing is intermarket-analysis: the bond-fund trade is read through conditions in the mining-stock index rather than in isolation.

The first rule is an index-jump-signal, a short-window percentage jump in that index, with the lookback generally placed in a five-to-25-day range. The premise is that an abrupt index move is followed by a same-direction move in long-term rates and an opposite move in bond-fund prices.

A price-confirmation-filter also has to be true. A buy requires the bond fund to have risen more than a set percentage from the lowest low recorded while the stance was short. A sell requires the fund to have fallen more than a set percentage from the highest high recorded while the stance was long.

Both the index-jump event and the bond-fund confirmation must be true on the same day, and the stance change is specified to occur on that day's close. Those clauses make one mechanical-trading-system: specified inputs and market state become a long or short stance under a stated close-to-close execution constraint.

Why the daily index version exists

A predecessor weekly gold-mining series could not act until the weekend. That left a one-week action gap that the daily-index version was built to close.

How the walk-forward cycles were built

With only about 12 years of daily index history, a single estimation-versus-holdout split was rejected because the estimation half would be too short to cover many price-dynamic regimes. Walk-forward windows were used instead.

Each walk-forward cycle estimated the five parameters on a six-year window, applied those values to the next one-year holdout, then advanced the window by one year. The holdouts were merged into one out-of-sample-holdout evaluation segment. That is walk-forward-analysis: fit the rules on a moving in-sample window, apply those frozen values to the next unused stretch, advance the window, and judge the merged holdouts rather than one curve-fit.

To limit computation, the search ran in two passes. First came the index-jump thresholds and lookback without confirmation. Then the bond-fund confirmation percentages were estimated on a short list of first-pass survivors.

Unstable coefficients and a later break

Chosen parameter sets changed from one six-year window to the next. That change was treated as evidence that six years did not yield coefficients that stayed stable or captured all observed dynamics. It is a parameter-stability finding: a shift to the next estimation window did not leave the chosen values settled.

A later stretch in which the index and bond prices moved opposite to the assumed linkage was treated as an unavoidable rule failure. That stretch is market-decoupling: the assumed gold-stock and bond-price linkage fails, so a rule that is internally consistent can still produce a large, unavoidable loss. A three-to-four-percent protective stop was proposed as a later design addition.