如何用ggplot在真实时间序列上绘制LSTM股票预测值？

The issue with your current ggplot code is that it aligns all prediction sequences to the start of the x-axis instead of shifting them to their correct positions (like the Python code does with padding). Here's how to fix it by properly positioning each prediction segment:

Step 1: Prepare the Data Correctly #

We need to create a combined data frame where:

• The true test data is a continuous sequence from x=1 to x=length(y_test)
• Each prediction sequence is shifted to start at its correct position (i*prediction_len + 1 for the i-th sequence)
• We add a group identifier to ensure ggplot draws each prediction as a separate line segment (not connecting them to each other)

library(dplyr)
library(ggplot2)
library(purrr) # For map_dfr function

# Define prediction length (matches your function parameter)
prediction_len <- 50

# Create true test data frame
true_data <- tibble(
  x = 1:length(y_test),
  value = unlist(y_test),
  series_type = "True Data",
  group_id = "True_Data"
)

# Create prediction data frame with correct positioning
prediction_data <- purrr::map_dfr(seq_along(predictions), function(seq_idx) {
  # Calculate starting position for this prediction sequence
  start_x <- (seq_idx - 1) * prediction_len + 1
  end_x <- start_x + prediction_len - 1
  
  tibble(
    x = start_x:end_x,
    value = predictions[[seq_idx]],
    series_type = "Prediction",
    group_id = paste0("Pred_", seq_idx)
  )
})

# Combine true and prediction data
plot_data <- dplyr::bind_rows(true_data, prediction_data)

Step 2: Plot with ggplot2 #

Now we can plot the data, using series_type for color and group_id to ensure proper line segments:

ggplot(plot_data, aes(x = x, y = value, color = series_type, group = group_id)) +
  geom_line(linewidth = 0.8) +
  labs(
    title = "S&P 500 Normalized Price: Predictions vs True Data",
    x = "Time Step",
    y = "Normalized Closing Price",
    color = "Series"
  ) +
  theme_minimal() +
  theme(plot.title = element_text(hjust = 0.5))

Why Your Original Code Failed #

Your original approach using cbind.fill and gather aligned all prediction sequences to the first 50 x-values, causing all predictions to overlap at the start of the plot. By explicitly calculating the correct x-position for each prediction sequence and adding a group identifier, we replicate the Python code's behavior where each prediction shifts forward by prediction_len steps.

This will produce a plot identical to the Python version: the true data is a single continuous line, and each prediction segment starts at the correct position in the time series.

内容的提问来源于stack exchange，提问作者Mislav