Concept

LLM/transformer-derived Signals

LLM/transformer-derived Signals are Machine Learning concepts. A reference entry: the Library explains it rather than implements it.

What are LLM/transformer-derived signals?

LLM and transformer-derived signals are trading inputs produced by transformer models, the neural architecture behind large language models. They come in two flavors. The first feeds text (headlines, earnings-call transcripts, filings, social posts) through a language model to score sentiment, classify events, or extract structured facts. The second applies transformers directly to numeric sequences: attention-based forecasters and time-series foundation models trained on large collections of historical series.

What separates transformers from earlier neural networks is attention: rather than processing a sequence step by step, the model learns to weight every part of the input against every other part. For text, that captures context that dictionary-based sentiment scoring misses. For price series, it lets a model reference distant history without squeezing it through a compressed running memory.

Trading applications treat the model as an upstream data producer, not a strategy. Raw outputs, whether a sentiment score, an event tag, or an embedding vector, are handled like any other product of feature engineering: combined, validated, and sized downstream, often via logistic calibration or ensemble voting. Published evidence points to modest, fast-decaying predictive content at short horizons, not large standalone edges.

Why there's no indicator for this

A chart indicator computes bar by bar from a symbol's price and volume history; a transformer signal cannot be produced that way. The model is a file of millions to billions of learned parameters that needs real inference hardware, and its interesting inputs, such as news wires, transcripts, filings, and social feeds, never appear in OHLCV data. The most capable models sit behind paid APIs, and even open-weight alternatives are far too large for a chart script to embed or recompute.

Vendors do sell LLM-derived sentiment and event feeds, and a feed can be plotted as an external series. Know what that proxy is: a snapshot of one vendor's model, prompts, and training cutoff. It can capture aggregate tone around a symbol; it cannot be audited from the chart, reproduced independently, or assumed stable as the vendor updates models. Any overlay claiming to run an LLM on price data alone is mislabeled.

How to evaluate an LLM signal feed

You cannot recompute a vendor's model, but you can audit its discipline before trusting the output.

1Demand point-in-time data: every score should carry the timestamp when it was actually available, not a later revision or re-run.
2Check for training-cutoff leakage: a model trained on data that overlaps the backtest window has effectively seen the answers.
3Calibrate scores against realized outcomes on your own horizon; a raw score is not a probability until it has been tested.
4Track live performance against the pitch, since vendors can update underlying models without notice and behavior drifts.

How traders use it

Scoring earnings calls and news: desks run transcripts and headlines through language models to flag tone shifts versus prior quarters, then read the price reaction against the score rather than trading it alone.
Event classification: tagging headlines by type (guidance cut, share offering, litigation, M&A) to route symbols into pre-defined playbooks faster than manual reading allows.
Feature generation: text embeddings and topic scores become candidate model inputs, subject to the same label and horizon discipline as any learned feature.
Forecasting research: quants benchmark attention-based forecasters against simpler baselines such as kernel regression or a random forest; the heavy model earns a seat only if the gap survives out-of-sample testing with costs.

Related concepts · Learned models

Kernel Regression6 Neural Networks3 Gaussian Process Regression2 Kernel Density Estimation2 Support Vector Machines1 Random Forest1 Bayesian Classifiers1 Self-organizing Maps1 Reinforcement-learning Policies1 LSTM / Recurrent Networks0

Concept family

Machine Learning

32 concepts mapped · 21 in the Library

LLM/transformer-derived Signals FAQ

Do LLM signals actually predict markets?

Studies find modest short-horizon predictive content in text-derived scores, concentrated around news events and decaying within hours to days. Nothing published supports large, stable, standalone edges, and transaction costs erode much of what exists.

Can a chart script run a transformer?

Not honestly. Chart scripts lack the compute, the parameter storage, and above all the text data. Indicators marketed as transformer-powered are either plotting an externally produced feed or using the label loosely.

What data do these signals consume?

News wires, transcripts, regulatory filings, and social media for language models; long panels of historical series for time-series transformers. All of it lives outside the chart.

Are transformer forecasters better than classical models?

Mixed. Time-series foundation models show promise on some benchmarks, but on noisy financial series simple baselines remain hard to beat once judged strictly out of sample.

Build LLM/transformer-derived Signals your way.

Quant writes, tests, and refines it with you — then it runs on LuxAlgo charting or ports to TradingView.