Walk-forward screen for yen indicator rules

Indicator rules on yen futures

A moving-average is a lookback smoother of ordered prices used as a parameterized forecast rule whose settings are locked on one sample and scored on the next. A relative-strength-index is a bounded oscillator of sequential price changes used as a parameterized signal in the same walk-forward screen. A stochastic-oscillator is a range-position oscillator on ordered highs, lows, and closes used as a parameterized signal in the same walk-forward screen.

The evaluated battery comprised 34 parameterized rules on exchange-listed yen futures and included those three model families. Each rule was run with one to three parameters swept over a grid.

Walk-forward lock

A walk-forward-lock chooses parameters on one year of nearby-contract data, trades only those settings on the following unused quarter, then rolls both windows forward. In this historical workflow, parameters were locked on one year of nearby-contract data, applied only to the next unused quarter, then both windows were rolled forward one quarter. That procedure produced out-of-sample results from March 2, 1978 through September 1, 1987.

Stances and position modes

Twenty-eight of the 34 rules were coded as buy-strength-sell-weakness trend followers. That stance buys a rising market and sells a falling one. The rest were coded as buy-weakness-sell-strength counter-trend rules, which buy a falling market and sell a rising one.

Three position modes were scored separately. Speculative-reversal holds a long or short futures position until the next opposite signal replaces it. One cash-hedge-mode takes only sell signals against a cash long and goes flat on buys. The other cash-hedge-mode takes only buy signals against a cash short and goes flat on sells.

Zero-mean test and autocorrelation filter

A one-tailed t-test that mean quarterly out-of-sample net returns exceeded zero left 20 of 34 rules in two-sided speculative mode, 7 in exclusive-short mode, and 2 in exclusive-long mode.

An autocorrelation-filter drops a rule if its quarterly net-return series shows significant first-order serial correlation that would inflate a t-test. After that filter, 16 speculative rules, 5 short-hedge rules, and 1 long-hedge rule remained.

Pairwise check among survivors

Among the rules that survived the autocorrelation filter, pairwise tests found no statistically superior mean net quarterly return.

Cost, interest, and distribution assumptions

Quarterly results assumed one contract, a $47.50 cost made of a $10 commission plus three ticks of slippage, and interest on a $3,500 margin at the matching 90-day Treasury-bill rate, with no interest credit on equity above exchange minimum margin.

The t-test assumed normal, serially independent quarterly net returns. A Kolmogorov-Smirnov check did not reject normality, so first-order correlation was used as the separate independence screen.