"The Performance of the LSTM-based Code Generated by Large Language Models (LLMs) in Forecasting Time Series Data"

This paper evaluates how LLMs can generate LSTM code for time series forecasting, helping data analysts automate deep learning model creation.

-----

https://arxiv.org/abs/2411.18731

Original Problem 🤔:

Data analysts need LSTM models for time series analysis but lack coding expertise. Manual implementation requires deep understanding of neural architectures, making it inaccessible for many practitioners.

-----

Solution in this Paper 🛠️:

→ Researchers tested GPT-3.5, PaLM, Llama-2, and Falcon on financial time series forecasting tasks.

→ They designed prompts with varying levels of clarity, objectives, context, and format.

→ Generated LSTM code was compared against manually optimized implementations.

→ Temperature and configuration parameters were systematically analyzed.

-----

Key Insights 💡:

→ Simple prompts often outperformed complex ones

→ Lower temperature settings (0.1) reduced hallucination compared to higher values (0.7)

→ One or two LSTM layers with 50 nodes proved optimal

→ Batch size of 32 showed best results across models

-----

Results 📊:

→ GPT-3.5 achieved lowest RMSE on 8/10 datasets

→ Generated models outperformed manual implementations on 7/10 datasets

→ Temperature of 0.1 reduced invalid model generation by 40%