Time Series Forecasting with AI

Introduction

Time series forecasting is crucial for:

Sales and demand forecasting

Financial market prediction

Energy consumption planning

Predictive maintenance

Weather forecasting

Classical Methods

python
from statsmodels.tsa.arima.model import ARIMA
import pandas as pd
Load time series data
ts = pd.read_csv('sales.csv', index_col='date', parse_dates=True)
Fit ARIMA model
model = ARIMA(ts, order=(2, 1, 2))  # (p, d, q)
results = model.fit()
forecast = results.forecast(steps=30)

LSTM-Based Forecasting

python
import numpy as np
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Dropout
def create_sequences(data, seq_length):
    X, y = [], []
    for i in range(len(data) - seq_length):
        X.append(data[i:i+seq_length])
        y.append(data[i+seq_length])
    return np.array(X), np.array(y)
model = Sequential([
    LSTM(128, return_sequences=True, input_shape=(seq_length, n_features)),
    Dropout(0.2),
    LSTM(64),
    Dropout(0.2),
    Dense(1)
])
model.compile(optimizer='adam', loss='mse')

Transformer-Based: Temporal Fusion Transformer

python
from pytorch_forecasting import TemporalFusionTransformer, TimeSeriesDataSet
dataset = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="value",
    group_ids=["series_id"],
    max_encoder_length=60,
    max_prediction_length=24
)
tft = TemporalFusionTransformer.from_dataset(dataset)
trainer = pl.Trainer(max_epochs=30)
trainer.fit(tft, train_dataloader, val_dataloader)

Foundation Models for Time Series

Amazon Chronos

python
import torch
from chronos import ChronosPipeline
pipeline = ChronosPipeline.from_pretrained(
    "amazon/chronos-t5-large",
    device_map="cuda",
    torch_dtype=torch.bfloat16
)
Zero-shot forecasting - no training needed!
forecast = pipeline.predict(
    context=torch.tensor(historical_data),
    prediction_length=24,
    num_samples=20
)

Nixtla TimeGPT

python
from nixtla import NixtlaClient
client = NixtlaClient(api_key="...")
forecast = client.forecast(df, h=24, freq="H")

Model Selection Guide

Data SizeBest Approach

< 1000 pointsARIMA, Exponential Smoothing 1000-100KLightGBM with time features > 100KLSTM or TFT Multiple seriesFoundation model (Chronos)