Glossary
Plain-English definitions of the neural-network and signal-processing terms used across the site.
- AFL — AmiBroker Formula Language
- AmiBroker's built-in language for indicators, scans and trading systems. The Toolbox plugs into AFL, so every scanner, indicator and neural network is called as an ordinary AFL function — there's no separate app to drive.
- MLP / feed-forward network
- The standard neural network: inputs feed forward through one or more hidden layers to an output, with no memory of earlier bars. Fast and robust — the Toolbox's default network, and the only kind that can be compiled to pure AFL.
- LSTM — Long Short-Term Memory
- A recurrent network with a gated memory cell that reads a window of consecutive bars in order, so it can learn patterns that depend on the recent sequence of price, not just the latest bar.
- GRU — Gated Recurrent Unit
- A streamlined recurrent network — a lighter cousin of the LSTM with similar sequence-memory but fewer parameters, so it often trains faster.
- Walk-forward
- Repeatedly retraining a model on the bars immediately before each bar and predicting ahead, so every prediction is made from past data only. A realistic, out-of-sample picture of how a model would have behaved in real time.
- Out-of-sample
- Data the model was not trained on. Performance out-of-sample is the honest measure; strong results only on the data a model was trained on usually mean overfitting.
- Overfitting
- When a model learns the noise in the training data instead of the signal — so it looks brilliant on the bars it was trained on and falls apart on new ones. The real challenge on noisy market data.
- Dropout, weight decay & early stopping
- Techniques that fight overfitting by limiting how tightly a network can fit the training data: randomly dropping connections during training, penalising large weights, and stopping before the network starts to memorise.
- DFT — Discrete Fourier Transform
- A way to break a series into the cycles (frequencies) that make it up. The basis of the Toolbox's Goertzel cycle indicator.
- FFT — Fast Fourier Transform
- A fast algorithm for the DFT. The Toolbox's end-point FFT uses it to filter noise out of price without looking into the future.
- Goertzel algorithm
- An efficient way to measure specific frequencies in a series — used in signal processing, and famously in touch-tone phones. The Toolbox applies it to find the dominant market cycles.
- Hilbert transform
- A signal-processing tool for measuring the phase and dominant cycle of a series; the basis of several of the Ehlers cycle indicators in the Toolbox.
- Homodyne discriminator
- John Ehlers' classic method for reading the dominant cycle: it treats price like a radio signal, splits it into in-phase and quadrature components, and reads how fast the phase is rotating. Low-lag and smooth — the method the original public versions of the cycle indicators used, and still the Toolbox's default.
- DSP engine / dominant-cycle estimator
- The interchangeable piece inside the engine-aware cycle indicators that measures the dominant cycle. Ehlers' originals were hard-wired to the homodyne discriminator; the Toolbox lets you swap in other estimators — the autocorrelation periodogram (Mesa), the Burg maximum-entropy spectrum, a Kalman cycle tracker and a concentrated-taper (Multitaper) periodogram — to trade smoothness, resolution and lag without rewriting your formula.
- DSP — Digital Signal Processing
- The branch of maths behind filtering and cycle analysis. The Toolbox's cycle tools (Goertzel, FFT, MAMA, Hilbert) apply DSP techniques to price.
- Spread support
- The indicators work on a spread between two symbols (e.g. a pairs trade), not just a single instrument — true of every indicator except pattern exploration.
Back to all features.