Inference behavior on Daily Macro Data
Hi there, thanks for releasing the Kronos architecture! I've been testing Kronos-small on historical equity data.
When feeding it raw daily data (capturing a macro trend scaling from roughly 750 to 2700 over ~400 lookback steps), the model tends to output a flat, dampened prediction around the 1500–1750 level, missing the upward breakout. I tried applying a standard Min-Max normalization to the price columns before inference, but the forecast still tracks relatively flat.
Given that the model was primarily trained on high-frequency (5-minute) intervals, does it require a specific standardization/scaling technique for daily macro data, or is the architecture strictly optimized for Intraday volatility structures?
Stock used: MCX NSE Time Frame: 2 years
Any updates on this please???