We propose novel HARQ prediction schemes for Cloud RANs (C-RANs) that use feedback over a rate-limited feedback channel (4 and 8 bits) from the Remote Radio Heads (RRHs) to predict at the User Equipment (UE) the decoding outcome at the BaseBand Unit (BBU) ahead of actual decoding. In particular, we propose a novel dual-input denoising autoencoder that is trained in a joint end-to-end fashion over the whole C-RAN setup. In realistic link-level simulations at 100 GHz in the sub-THz band, we show that a combination of the novel dual-input denoising autoencoder and state-of-the-art SNR-based HARQ feedback prediction achieves the overall best performance in all scenarios compared to other proposed and state-of-the-art single prediction schemes. At very low target error rates down to $1.6 \cdot 10^{-5}$, this combined approach reduces the number of required transmission rounds by up to 50\% compared to always transmitting all redundancy.