This paper investigates wavelet-based preprocessing for currency exchange rate forecasting within an adaptive neuro-fuzzy inference system (ANFIS). While the integration of discrete wavelet transform (DWT) with ANFIS has been explored in earlier studies, our contribution lies in systematically identifying the most effective wavelet family and decomposition level for this task. Using 1-minute mid-price data (the mean of bid and ask) of the Polish zloty (PLN) against USD, CAD, CHF, JPY, and EUR, we evaluate seven wavelet families across three decomposition levels (2–4). Results indicate that the reverse biorthogonal wavelet (rbio2.4) at level 3 yields the best predictive performance, providing a balance between capturing long-term trends and short-term fluctuations. Compared to models trained directly on raw data, the optimized wavelet-enhanced ANFIS reduces the effective training set size to approximately 12.5% of the original while significantly lowering forecast errors and improving stability. These findings highlight the importance of wavelet selection as a critical design choice in hybrid financial time series forecasting models.