Constructing stochastic, RSI and CCI inputs for forecasts

What a learner can actually weight

TradersWeek editorial reading: treat stochastic, RSI and CCI construction as the first forecast-design decision. A time-series learner can weight only the scale, cycle length and independence you encode. It cannot apply the rule sheet a discretionary reader would draw on the same chart.

A forecast model assigns input weights by comparing each produced output with a supplied training-pattern target. The training-pattern is the known result paired with that input row. If the error count does not fall, the columns may be complete as an analyst worksheet and still unusable as model inputs.

When columns move together

Currency-pair, Eurorate and gold columns tested against Deutschemark per US dollar were highly correlated and showed little lag. That input correlation left little independent information, so raw prices alone were treated as low-value forecast inputs. Outputs from separately trained models were then fed into a primary model.

Fixed and cycle-linked stochastic construction

A stochastic oscillator is a bounded oscillator from ordered prices over a stated lookback and smoothing window. A cotton stochastic construction uses an eight-period lookback and three-period smoothing and still shows a varying lag.

A companion construction multiplies that fixed period by a sinewave so oscillator length tracks a selected dominant cycle. That offset is a variable-length displacement: the period lengthens or shortens as the selected dominant cycle expands or contracts.

Band encodings for RSI and CCI

A relative strength index is a momentum oscillator over a stated lookback. A commodity channel index is a deviation oscillator over a stated lookback. Encoding either series as the arithmetic gap to a volatility band can invert importance, because the model treats larger numbers as more significant while that gap shrinks as the oscillator approaches the band. A band touch is required to outrank a mid-band reading. The volatility band is an envelope around the oscillator used as that reference level.

Weekly Deutschemark constructions place a 14-period RSI and a 23-period CCI each against a volatility band. Separate bull-trend and bear-trend volatility columns, plus inverse and four-period-displaced smoothed RSI encodings, are listed as alternative input tests.

Checks, row counts and staged construction

When in-sample error looks resolved but held-out rows fail, the model may have memorized pairs. Multiplying a training-pattern column by minus one, reducing hidden units, or randomizing chronological rows is presented as a check against that shortcut. Row randomization reorders chronological training examples so a trending series cannot be generalized from presentation order alone.

A market-forecasting guideline is fifteen example rows per input column, so four columns imply sixty rows. The model sees one row at a time, so any prior-week price must be written into that same row.

Construction is described as training on one indicator first and then adding columns, rather than loading every available series at once. Accepted training tolerances are widened early and narrowed in later stages.