Evaluating weekly intermarket context for equity regimes

What the four inputs were asked to do

The archive workflow used intermarket context, meaning weekly changes in short-term yields, long-bond yields, gold, and a dollar index, to place an equity-index stance in a cross-market regime rather than in a single-market chart.

The four inputs were weekly changes in the three-month bill rate, the 30-year bond yield, gold, and a dollar index. Those series were used to forecast the next week's equity-index change.

The network used a single hidden layer, an intermediate mixing stage between the cross-market inputs and the next-week equity-index output, with an option to add one hidden neuron every 10 training runs.

How much history the mapping was given

A sample-length heuristic set a working range of a few hundred multiples of the forecast horizon so the history would span mixed conditions without leaning on obsolete links. The working rule was to analyze about 200 to 300 times the forecast horizon, so a one-week forecast would rest on 200 to 300 weeks of history.

A second working rule, the input-to-history ratio, kept the number of analyzed periods at least 20 times the number of input series, or about twenty history periods per input series. Near-duplicate inputs, closely related series that add little independent information, were to be dropped so the mapping would not lean on isolated sample quirks or memorize sample noise.

Holdout check and error-driven retuning

The learning window ran 257 weeks from 31 December 1986 through 27 November 1991. A holdout check withheld 20 percent of those weeks, 51 weeks, from weight updates so right-versus-wrong rates could be compared with the weeks used for learning. A further 29 weeks from 4 December 1991 through 17 June 1992 sat outside both learning and testing.

When a forecast fell outside a user-set tolerance of the next week's actual equity-index change, error-driven retuning sent that out-of-tolerance forecast error backward through the connection weights so the network could adjust.

When later passes became a memorization signal

In more than 650 training-and-testing runs, holdout correctness often tracked training correctness between runs 180 and 280. After run 280, holdout correctness trended down while training correctness held or rose. That pattern was taken as a memorization signal: a shift from relationship learning to memorizing sample nuances.

The network from run 200 was applied to the full weekly file through 17 June 1992, including the 29 weeks that had been excluded from training and testing.

Why full convergence was not the goal

Because an equity index is driven by more influences than the four inputs and those influences shift, the work treated the task as unlike training on a fixed natural law that can be forced to full convergence.

Network opacity remained even when the holdout check looked stable: the trained mapping cannot be unpacked into an explicit rule set. Editorial note: a mapping that still looks tidy on trained weeks after many later passes is not, by itself, a sign of a usable regime map.