The previous post built a vanilla Recurrent Neural Network — a model with a hidden state that carries forward bar to bar — and then ran straight into its defining limitation. Training a recurrent network means multiplying the gradient by a tanh derivative at every step back through time, and a long chain of numbers…
Standard SuperTrend gives you a clean trailing stop in trends and brutal whipsaws in noise. LuxAlgo’s Evasive SuperTrend is built for the noise problem — when price gets within a configurable Threshold × ATR of the active band, the band actively pushes away from price by Alpha × ATR, dodging the wick that would have…
When you call AddPlot, NinjaScript creates a Series<double> behind the scenes that you access through Values[i]. That’s a custom Series with nothing special done to it — it’s the same type you can create for your own use when you need per-bar tracking that doesn’t belong on a plot. This post covers when to reach…
Every model in this series so far — k-Nearest Neighbors, online logistic regression, the single– and multi-hidden-layer nets, and last post’s Adam-trained version — has shared one quiet limitation. Each prediction sees only the current bar’s feature vector. If a pattern depends on what the market did five bars ago, that information has to be…
Bill Williams fractals — five-bar pivots where the middle bar’s high is highest (or low is lowest) — are a clean structural primitive but useless on their own as a trade signal. BigBeluga’s Fractals Trend takes the next step: store the most recent N fractal highs and N fractal lows, smooth them into trend bands,…
An SMA cross that fires at 3 a.m. on overnight NQ is rarely useful. A reversal setup at 11:58 during the lunch chop is worse than useless. Most trading logic belongs in specific windows, and NinjaScript has simple tools for keeping it there — a couple of integer comparisons and a helper method you write…
The prior post let you stack any number of hidden layers — but every weight update was still plain stochastic gradient descent: one update per bar, one fixed step size for every weight. Plain SGD is the simplest possible training rule, and it works on shallow networks, but two things tend to go wrong at…
capissimo’s Chop and explode (ps5) reframes the classic RSI as a regime detector. Instead of reading 70/30 thresholds for overbought/oversold, this version applies RSI to a minimax-scaled source — the price is normalized to a 0-100 range over a window — and uses 60/40 as the BUY/SELL switches. When the scaled RSI sits between 40…
Tooltips live in two places in a NinjaScript indicator. The first is the settings dialog — the small hover text that explains what a property does. The second is the chart itself — labels and annotations attached to drawn objects that show context when a trader hovers over them. Both are cheap to add, both…
The prior post built a single-hidden-layer neural network — three inputs, six tanh neurons in one hidden layer, one sigmoid output, 31 parameters. It worked because the hidden layer can compose the raw features into non-linear combinations that a single neuron cannot represent. The post closed with a hint: “add a hidden layer + backprop…
Notifications
