Cross validation protocol sensitivity in small sample financial deep learning for gold price classification

This study develops a pretrained visual representation pipeline for financial time series classification, applied to gold futures (GC = F) direction prediction, under a chronologically purged and embargoed cross validation protocol that avoids residual leakage a randomly assigned protocol can permit between non overlapping windows. Gold futures closing prices (2000–2025) are segmented into 207 non overlapping 30 day windows, rendered as three encodings (line, candlestick, Gramian Angular Field), and processed by pretrained visual encoders (DINOv2, ConvNeXt Tiny, CLIP) with balanced three class labels (SELL/HOLD/BUY). An exploratory encoder selection stage identifies the dual fusion of ConvNeXt Tiny and CLIP as the best of seven combinations (single split accuracy 73.8%; cross validated accuracy 64.0% ± 6.0%); DINOv2 is evaluated only at this stage. This study’s primary, confirmatory result comes from a single pass, five fold, purged and embargoed cross validation applied to six model and fusion variants: a raw OHLC baseline and three sequence native baselines (LSTM, GRU, a lightweight Transformer) remain indistinguishable from the 33.3% chance level, while the two pretrained visual representations retain a smaller but positive signal (ConvNeXt+CLIP: 44.0% accuracy, MCC = 0.158; Multi Encoding Fusion: 42.0% accuracy, MCC = 0.129). The 44.0% figure, not the exploratory 64.0% estimate, is this study’s headline finding; a transaction cost backtest of the four non visual representations shows flat to mildly positive returns despite near chance accuracy. Pretrained visual representations retain a modest, above chance signal under a conservative protocol, underscoring that reporting cross validation protocol alongside headline accuracy is essential at this sample size.