Neural Network Wizard

Walk-Forward Networks

A Walk-Forward network is the adaptive alternative to a Standard one. Rather than training once and saving the result, it re-trains on a rolling window of recent bars as it walks forward through the data — so the model continually adapts to the most recent market behaviour. It is more adaptive but slower, works on a single symbol, and saves no network file.

You choose this mode at startup (or via File → New). In Walk-Forward mode several Standard-only controls are gated off, and one Walk-Forward-only control is enabled.

Standard vs. Walk-Forward: which to choose

The two modes answer the same question in opposite ways. A Standard network learns the market once and assumes that lesson keeps holding; a Walk-Forward network assumes the market drifts and keeps relearning from the most recent bars. Neither is strictly better — it's a trade-off:

 StandardWalk-Forward
TrainingOnce, then reuse the saved .netContinuously, on a rolling window
Adapts to new regimesNo (until you retrain)Yes, automatically
SymbolsA whole poolOne symbol only
SpeedFast to plot (pre-trained)Heavier — trains as it runs
ModelsMLP + recurrent LSTM/GRUFeed-Forward MLP only
Best forMost users; iterating and comparingA single instrument in a shifting market

If you're undecided, build a Standard network first — it's faster to iterate on and unlocks every feature. Reach for Walk-Forward when you're focused on one instrument and want the model to keep adapting on its own.

Gated controls in the main window

The main window in Walk-Forward mode, with several controls gated.
The main window in Walk-Forward mode.
  1. Generate AFL — disabled for Walk-Forward (network-to-AFL export is Standard + MLP only).
  2. Walk-Forward Lookback — ENABLED for Walk-Forward: recent bars the network re-trains on at each step. Default 200.
  3. Testing Data (%) — disabled for Walk-Forward (the rolling re-train has no fixed held-out split).
Note

Walk-Forward mode is Feed-Forward (MLP) only and single-symbol. The recurrent LSTM/GRU models, the ticker training pool, the fixed test split and the network-to-AFL export all belong to Standard networks and are unavailable here.

Tip

The Walk-Forward Lookback is the heart of this mode: it sets how many recent bars the network re-trains on at each step. A shorter lookback adapts faster to regime changes but sees less data; a longer one is steadier but slower to react.

The generated formula (Walk-Forward)

Because a Walk-Forward network trains and trades in one continuous pass, the Wizard produces a single AFL formula rather than separate training and indicator formulas.

The Generate Formula window for a Walk-Forward network, showing a single formula.
The single Walk-Forward formula.
  1. Neural Network Formula — the single Walk-Forward AFL that re-trains on a rolling window and trades in one pass.
  2. OK — close the window; copy the formula into an AmiBroker chart.
Warning

A Walk-Forward formula does its training every time it runs, on the chart it is applied to. That makes it noticeably heavier to compute than a pre-trained Standard indicator — expect slower chart refreshes, especially with a large lookback or many bars.