Skip to main content
Time-LLM is a reprogramming framework to repurpose LLMs for general time series forecasting with the backbone language models kept intact. In other words, it transforms a forecasting task into a “language task” that can be tackled by an off-the-shelf LLM. References Figure 1. Time-LLM Architecture. Figure 1. Time-LLM Architecture.

​
1. Time-LLM

​
Usage example

import pandas as pd
import matplotlib.pyplot as plt

from neuralforecast import NeuralForecast
from neuralforecast.models import TimeLLM
from neuralforecast.utils import AirPassengersPanel

Y_train_df = AirPassengersPanel[AirPassengersPanel.ds<AirPassengersPanel['ds'].values[-12]] # 132 train
Y_test_df = AirPassengersPanel[AirPassengersPanel.ds>=AirPassengersPanel['ds'].values[-12]].reset_index(drop=True) # 12 test

prompt_prefix = "The dataset contains data on monthly air passengers. There is a yearly seasonality"

timellm = TimeLLM(h=12,
                 input_size=36,
                 llm='openai-community/gpt2',
                 prompt_prefix=prompt_prefix,
                 batch_size=16,
                 valid_batch_size=16,
                 windows_batch_size=16)

nf = NeuralForecast(
    models=[timellm],
    freq='ME'
)

nf.fit(df=Y_train_df, val_size=12)
forecasts = nf.predict(futr_df=Y_test_df)

​
2. Auxiliary Functions

​
ReprogrammingLayer

ReprogrammingLayer(
    d_model, n_heads, d_keys=None, d_llm=None, attention_dropout=0.1
)
Bases: Module ReprogrammingLayer

​
FlattenHead

FlattenHead(n_vars, nf, target_window, head_dropout=0)
Bases: Module FlattenHead

​
PatchEmbedding

PatchEmbedding(d_model, patch_len, stride, dropout)
Bases: Module PatchEmbedding

​
TokenEmbedding

TokenEmbedding(c_in, d_model)
Bases: Module TokenEmbedding

​
ReplicationPad1d

ReplicationPad1d(padding)
Bases: Module ReplicationPad1d