Signal Analysis, Models & Validation
This is the validation layer: everything between writing a Signal and committing to a backtest. It covers the Signal tab and its time-series chart, the four EDA charts that live inside that tab, benchmark context, anchor plans, return datasets, tunable parameters, and the Model stage that combines Signals into one prediction. The section is a map. Each page describes one surface in depth.
Where this fits in the pipeline
The shape of the research pipeline is Signal (and its EDA) → Model → Backtest. A Signal reads market data and produces a value at every point in time. Its EDA charts hang off the Signal itself, not a separate screen. A Model is the optional next rung: it weights several Signals into one predicted value or class probability. A backtest is the terminal rung, and it is not optional. Every workflow ends in a backtest; the stages before it run only when the strategy needs that much validation.
Signal evaluation, save, extend, and import compute over one symbol and a date range through the Signal language. The visual layer on top covers the Signal tab chrome, the four EDA charts, benchmark context and the inheritance badge, anchor plans, return datasets, tunable parameters, and the entire Model stage. Each page below describes one surface in definite terms.
Two facts bind every page here. The accessor spine runs 09:31 through the session close (390 observations on a normal day, 210 on a half day). And bar.* prices are split- and dividend-adjusted by default, so results in this section read against adjusted bars. For where those come from, see data accessors and stock price adjustments.
The pages in this section
- The validation layer, from Signal to Model places the stage in the pipeline and shows how the rungs connect.
- The Signal tab and the time-series chart is the entry point: write a Signal, evaluate it over a symbol and date range, and read its value series on the chart.
- Reading the EDA charts covers the four charts (Time Series, Distribution, Scatter, Decile) and what each one tells you about whether a signal predicts trade outcomes.
- Benchmark context and the inheritance badge sets the outcome axis the EDA charts read against, and shows where that setting came from.
- Anchor plans fix how comparison observations are sampled when the benchmark is another signal rather than a return dataset.
- Return datasets record the historical outcomes of a fixed trade structure and are reused three ways: the outcome side of an EDA benchmark, an auto-fill source for a backtest position, and the target a model learns to predict.
- Tunable parameters expose a value as a named knob, with a manual range sweep that runs one backtest per value and overlays the results.
- Building and training a model walks the Model stage: pick features from saved signals, choose a target, pick one of four model types, and train under walk-forward validation.
- Model outputs and validation reads a trained model's family-specific views, all built to answer one question: does it predict outcomes on data it never saw.
- Using a model in a backtest wires a validated prediction into entry, exit, or sizing at the terminal stage.
Where this connects
Back to the Signal language for the DSL that produces the value series these charts read, and to Data & Methodology for the accessors and the defaults that shape every number. Forward to Backtests, the terminal stage every workflow in this section feeds.