Using LLMs for Cryptocurrency Trade Signal Generation

Large language models (LLMs) can generate structured cryptocurrency trade signals, but valid commands and stable decisions do not necessarily produce profitable trades. We evaluate an implemented signal-generation pipeline through 4,663 model-case-condition evaluations across six experiment families: repeated-call consistency, input representation and model comparisons, setup-selected market cases, input modality, prompt style, and model-specific knowledge-cutoff groups. Models receive historical price and volume tables, indicator-enriched representations, market-structure summaries, or chart images. Signals are validated and evaluated with a common historical backtester incorporating KuCoin contract trading fees and historical funding. Repeated GPT-5.5 calls often converge on stable directions and price levels, yet within-model and cross-model consensus do not consistently predict higher net profit and loss. Representation effects depend on model identity rather than following a universal advantage from richer context. Hybrid text-and-chart input leads the GPT-5.5 modality comparison, while visual-only input lags. Smart Money Concept instructions strongly alter analytical vocabulary and reshape the payoff distribution, but do not establish a profitability advantage. Temporal comparisons show no simple post-cutoff deterioration. The main performance contrasts remain exploratory and statistically uncertain. These findings characterize LLM behavior within a reproducible trading interface while identifying limits from signal selectivity, model provenance, and simplified execution; they do not establish robust live-trading profitability.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-20
DOI
https://doi.org/10.21203/rs.3.rs-11375846/v1
Primary Topic
Stock Market Forecasting Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Using LLMs for Cryptocurrency Trade Signal Generation

Florian Schneider, Visieu Lac, Jonas Delpho
Research Square
Stock Market Forecasting Methods
preprint

Using LLMs for Cryptocurrency Trade Signal Generation

Florian Schneider, Visieu Lac, Jonas Delpho
preprint en

Abstract

Large language models (LLMs) can generate structured cryptocurrency trade signals, but valid commands and stable decisions do not necessarily produce profitable trades. We evaluate an implemented signal-generation pipeline through 4,663 model-case-condition evaluations across six experiment families: repeated-call consistency, input representation and model comparisons, setup-selected market cases, input modality, prompt style, and model-specific knowledge-cutoff groups. Models receive historical price and volume tables, indicator-enriched representations, market-structure summaries, or chart images. Signals are validated and evaluated with a common historical backtester incorporating KuCoin contract trading fees and historical funding. Repeated GPT-5.5 calls often converge on stable directions and price levels, yet within-model and cross-model consensus do not consistently predict higher net profit and loss. Representation effects depend on model identity rather than following a universal advantage from richer context. Hybrid text-and-chart input leads the GPT-5.5 modality comparison, while visual-only input lags. Smart Money Concept instructions strongly alter analytical vocabulary and reshape the payoff distribution, but do not establish a profitability advantage. Temporal comparisons show no simple post-cutoff deterioration. The main performance contrasts remain exploratory and statistically uncertain. These findings characterize LLM behavior within a reproducible trading interface while identifying limits from signal selectivity, model provenance, and simplified execution; they do not establish robust live-trading profitability.

Research Square
IU International University of Applied Sciences (DE)
Stock Market Forecasting Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon