Wavelet analysis+forecastBitcoinCRYPTO:BTCUSDMarketSpectrHi Traders, I’ve been working on a quantitative forecasting model for crypto (BTC, ETH) and I’m planning to bring it live soon. Before that, I’d like to get some feedback from the community. The model applies wavelet decomposition to price data and uses that structure to generate forward price predictions. Instead of just predicting direction, it reconstructs an expected price path across a full horizon — from ~60 minutes up to 24 hours ahead. Each update produces a forward curve rather than a single point estimate. In other words: It captures multi-scale structure in the data (via wavelets) Projects that structure forward into the near future Starting today, I will try to publish a new 24-hour forecast every morning (Central European Time). Backtest (Bitcoin): The backtest was conducted in two modes: Mode 1: No fees (raw signal evaluation) Mode 2: With fees (~0.05%), only signals with strength > 1.013 are executed Evaluation metrics: Average directional accuracy across the full forecast horizon Pointwise cumulative directional accuracy (for each forecast step) Returns from simple long/short strategies Strategy setup: Long strategy initialized with $100 Short strategy with leverage equivalent to 0.001 BTC Results (All timestamps are in UTC) Mode 1 (no fees): Average directional accuracy approaches 0.518 toward the end of the backtest Pointwise cumulative accuracy, pointwise strategy return expectation, pointwise strategy return expectation with a weekly reset back to $100 and taking profit/loss. horizonaccuracyreturn,%weekly reset,% 60 min0.487107107 120min0.5148588 180min0.5187681 240min0.5228084 300min0.5207680 360min0.5214249 420min0.5185864 480min0.5186570 540min0.5186469 600min0.5186065 660min0.5204853 720min0.5195054 780min0.5195357 840min0.5214854 900min0.5194753 960min0.5176771 1020min0.5187075 1080min0.5185157 1140min0.5176165 1200min0.5176467 1260min0.5195256 1300min0.5193137 1360min0.5165863 1400min0.5148082 average0.5176267 Mode 2 (with fees + filtering): Average directional accuracy 0.513 Pointwise cumulative accuracy: Higher on shorter horizons Lower on longer horizons However: Even though longer horizons degrade in directional accuracy, the corresponding strategy returns are mostly not negative. This suggests the model still captures meaningful price movements, even when direction is harder to predict. Return expectations are clearly lower compared to Mode 1, but the strategy still survives transaction costs. Weekly reset gives some improvements to the return. horizonaccuracyreturn,%weekly reset,% 60 min0.5151918 120min0.573-14 180min0.5722828 240min0.5652524 300min0.565912 360min0.557-2-2 420min0.53599 480min0.5352020 540min0.52432 600min0.534-11-9 660min0.533-9-6 720min0.526-3-2 780min0.52923 840min0.5291211 900min0.523-6-6 960min0.515-11-10 1020min0.50712 1080min0.50048 1140min0.494-4-1 1200min0.49069 1260min0.49401 1300min0.49203 1360min0.4901216 1400min0.487610 average0.52446 Conclusion: The model captures a small but statistically consistent edge. As expected in a near-random market, performance depends heavily on execution and filtering - but the signal itself appears to contain usable structure. This is currently a pre-launch phase. Would this be useful to you? Would you consider paying for access? Any feedback is highly appreciated.