WiseTrader Toolbox 3.1

WiseTrader Toolbox 3.1 builds on the version 3 neural‑network engine with two new network types — a TCN and a Transformer — plus three ways to get more out of any network: ensembling, layer normalization, and categorical input embeddings. The recurrent networks also gained true multi‑layer depth, and the modern optimizer features (mini‑batch training and learning‑rate schedules) now apply to every network type.

Your existing formulas keep working with no changes: every new setting is optional and defaults to the previous behavior.

This is a free update for everyone with a version 3 license.

Highlights

• Two new network types — TCN & Transformer. Joining LSTM and GRU as sequence models that read a whole window of consecutive bars in order. The TCN (temporal convolutional network) is a fast, causal dilated convolution that reads the window in parallel; the Transformer uses multi‑head self‑attention to learn which past bars matter most. Reach for these when the sequence of recent bars carries the signal.
• Ensembling. Train several networks at once and average their predictions for a steadier, more reliable result. Works with every network type — the trained ensemble is saved and runs automatically, with no change to your run formula.
• Layer normalization. An optional stabilizer for deeper feed‑forward (MLP) networks that keeps training steady and often a little faster.
• Categorical inputs (embeddings). Feed a discrete input — a market regime, a day‑of‑week, or any integer category — the right way: the network learns a compact representation for each category instead of treating the ID as a number on a line. (Feed‑forward / MLP networks.)
• Deeper recurrent networks. The layers setting on LSTM and GRU now builds a genuinely stacked, multi‑layer network for greater capacity.
• Sequence models get the modern training features. Mini‑batch training and the learning‑rate schedules now apply to every network type, not just the MLP, and the per‑bar online‑training indicators work with the sequence models too.
• Neural Network Wizard, fully updated. The companion point‑and‑click wizard builds all of the above, validates your choices, and generates the matching AmiBroker formula for you.

Upgrading

Version 3.1 is a free update for every version 3 license holder — your existing version 3 key unlocks it. Install it over your current version and your license stays active.

Retrain your saved networks. Version 3.1 uses a new, unified on‑disk format for trained networks, so a .net file saved by an earlier build won't load — just retrain it once and you're set. Your formulas don't change.