Adapted from the QuantConnect Strategy Library write-up 'Combining Mean Reversion and Momentum in Forex Market' (after Serban, and Balvers & Wu). The source models log price as a random walk plus a stationary part, plus a momentum term: y_t = -(1-delta)*(x_{t-1}-mu)/sigma + rho*(x_{t-1}-x_{t-4}) + e_t. It fits delta and rho by OLS and trades the sign and size of the predicted next-period return, monthly, on 4 FX majors. FX is not tradable here, so we keep the model and change three things. (1) SINGLE INSTRUMENT, TIME-SERIES FORM. The source goes long the best-ranked pair and short the worst, but a 4-name ranking book is the multi-leg neutral family (L137) with a hand-picked universe (L151). Its own fallback rule is long-only if all predictions are positive and short-only if all are negative, and that is already a per-instrument sign trade, so we trade one instrument on the sign of its own prediction. (2) TIME COMPRESSION. Crypto cycles run roughly 4x faster than FX, and monthly rebalancing on ~6 years would give only ~70 decisions, which deflation always strips. The 3-month momentum lag becomes 21 days, the forecast horizon becomes 7 days, and the long-run mean mu and sigma come from a trailing 365-day window of daily log closes instead of the full-sample constant (the source's full-sample mean is look-ahead). The decision is made once per daily bar. (3) STRICTLY POINT-IN-TIME FIT. The source fits once on all history before the backtest. Here the OLS coefficients (intercept, b_rev, b_mom) are re-estimated each day on an EXPANDING window of past (features, realized 7-day forward return) pairs. Only pairs whose forward return has fully realized by the current bar are used, with a 180-sample burn-in. As in the source, samples with any |feature z| > 3 are dropped from the fit. Instrument selection, stated per L122: LTCUSDT.BINANCE was chosen because the model needs a long, continuous, liquid linear-perp history. It has multi-year daily data on Binance USD-M, while OKX's catalog currently holds only a few days of LTC/XRP bars, and Bybit and COIN-M give no advantage for a price-only signal (the COIN-M inverse valuation is currently flagged for engine review). It is not one of BTC/ETH/SOL, which dominate the corpus, and no supplementary feed is used (L160). We did not try other symbols first, and we pre-register LTC only. Flat state (L150): the strategy is flat whenever |predicted 7-day return| is below the deadband, and we expect it to be flat about 40-55% of days.