SkinCast
A multimodal Transformer sales forecasting model based on four-encoder fusion and global-local positional encoding. It can forecast day-by-day sales trajectories over 180 days for unreleased skins.
Timeline (last 90 days)
SkinCast is the core sales forecasting model in the SkinAdvisor framework, capable of predicting the day-by-day sales trajectory of unreleased skins over 180 days.
SkinCast contains four encoders: a pretrained Informer that processes player preference time series; a pretrained Player-Hero-Skin graph embedder that captures relational priors; and two lightweight Transformers that encode skin attributes and release context, respectively.
SkinCast's fused features pass through a shared backbone with local-plus-global day-x positional encoding, then through an N-layer Informer decoder that outputs the forecast sequence in a single non-autoregressive pass.
Non-autoregressive decoding requires only one forward pass at inference time, significantly reducing latency, and is one of the key designs enabling SkinCast to achieve millisecond-level response on CPU.
Offline experiments show that SkinCast reduces RMSE by about 20% compared with the strongest baseline, Informer, and by about 60% compared with classical methods.
Stress tests conducted on a 16-core CPU server confirmed that SkinCast's p99 latency is below 8 milliseconds and its peak throughput exceeds 1100 rps.
The paper points out that pretrained embeddings may leak temporal information, requiring a season-based split to prevent data leakage.
The SkinCast team cuts off temporal leakage through a season-based split, that is, training on data from earlier seasons and evaluating on data from later seasons.
All Facts (8)
SkinCast is the core sales forecasting model in the SkinAdvisor framework, capable of predicting the day-by-day sales trajectory of unreleased skins over 180 days.
50%UnverifiedSkinCast contains four encoders: a pretrained Informer that processes player preference time series; a pretrained Player-Hero-Skin graph embedder that captures relational priors; and two lightweight Transformers that encode skin attributes and release context, respectively.
50%UnverifiedSkinCast's fused features pass through a shared backbone with local-plus-global day-x positional encoding, then through an N-layer Informer decoder that outputs the forecast sequence in a single non-autoregressive pass.
50%UnverifiedNon-autoregressive decoding requires only one forward pass at inference time, significantly reducing latency, and is one of the key designs enabling SkinCast to achieve millisecond-level response on CPU.
50%UnverifiedOffline experiments show that SkinCast reduces RMSE by about 20% compared with the strongest baseline, Informer, and by about 60% compared with classical methods.
50%UnverifiedStress tests conducted on a 16-core CPU server confirmed that SkinCast's p99 latency is below 8 milliseconds and its peak throughput exceeds 1100 rps.
50%UnverifiedThe paper points out that pretrained embeddings may leak temporal information, requiring a season-based split to prevent data leakage.
50%UnverifiedThe SkinCast team cuts off temporal leakage through a season-based split, that is, training on data from earlier seasons and evaluating on data from later seasons.
50%