Package {GARCHInfoLSTM}

Type: Package
Title: GARCH-Informed LSTM Model for Volatility Forecasting
Version: 0.1.0
Description: The proposed Generalized Autoregressive Conditional Heteroskedasticity (GARCH)-informed Long Short-Term Memory (LSTM) model follows the concept of physics-informed machine learning (PIML) by integrating established econometric knowledge of price volatility into a data-driven forecasting framework. In the model, conditional volatility estimated from the GARCH process is incorporated as an additional explanatory signal or volatility-based weighting component within the LSTM architecture. This enables the LSTM to learn nonlinear temporal dependencies while remaining informed by the underlying characteristics of agricultural price series, including volatility clustering, heteroscedasticity and market uncertainty. The optimized weighting parameter, lambda, controls the contribution of the GARCH-derived volatility information to the final prediction. Thus, the model combines the statistical interpretability of GARCH with the nonlinear learning capability of LSTM, producing a hybrid PIML framework that is more responsive to both normal price movements and periods of extreme market volatility. The methodology is motivated by hybrid forecasting framework proposed by Yeasin and Paul (2024) <doi:10.1007/s11227-023-05542-3>.
Encoding: UTF-8
Imports: torch (≥ 0.11.0), rugarch (≥ 1.5.0), ggplot2 (≥ 3.4.0), cli (≥ 3.6.0), coro, stats, utils
License: MIT + file LICENSE
Config/roxygen2/version: 8.1.0
NeedsCompilation: no
Packaged: 2026-09-08 11:41:08 UTC; YEASIN
Author: Md Yeasin [aut], Ranjit Kumar Paul [aut, cre], Manojit Mandal [aut], Pushkar Bora [aut]
Maintainer: Ranjit Kumar Paul <ranjitstat@gmail.com>
Repository: CRAN
Date/Publication: 2026-09-21 20:50:08 UTC

Out-of-Sample Accuracy for GINN Volatility Forecasts

Description

Compares GINN conditional variance forecasts against realized variance computed from actual data that arrived after training. Returns RMSE, MAE, R-squared and SMAPE on variance units.

Usage

AccuracyGINN(fit, actual, lambda = NULL, verbose = TRUE)

Arguments

fit

A GINN object returned by GINN().

actual

Numeric vector of actual DATA observed after training. Their returns and realized variance are computed internally. Needs at least 2 values.

lambda

Lambda value to use. NULL = best lambda.

verbose

Logical. If TRUE (default), prints a formatted accuracy report to the console. Set to FALSE to suppress console output and only receive the returned data frame.

Value

Data frame with columns: Lambda, h, RMSE, MAE, R2, SMAPE (all in variance units). Also prints a formatted accuracy table to the console. The data frame has a "comparison" attribute with step-by-step Actual_Var vs Forecast_Var vs Error columns.

Examples


prices <- cumprod(1 + rnorm(60, 0.001, 0.02)) * 100
fit    <- GINN(prices, mode = "manual", lag = 5,
                ar_lag = 1, garch_p = 1, garch_q = 1,
                garch_mean = "constant", garch_dist = "norm",
                hidden_size = 4, num_layers = 1, lr = 0.01,
                epochs = 20, patience = 5,
                lambda_list = c(0.5), verbose = FALSE)

# Simulate actual future prices
new_prices <- tail(prices, 1) * cumprod(1 + rnorm(12, 0.001, 0.02))

acc <- AccuracyGINN(fit, actual = new_prices)
print(acc)

# Step-by-step comparison
attr(acc, "comparison")

GARCH-Informed Neural Network (GINN) – Volatility Forecasting

Description

Trains an LSTM on conditional variance sequences using the GINN composite loss: 

  Loss = (1-lambda) x MSE(sigma2_t, sigma2_hat_LSTM)
       + lambda     x MSE(sigma2_GARCH, sigma2_hat_LSTM)

Trains on FULL data – no train/test split. Best lambda selected by Training RMSE. Use predictGINN() for out-of-sample volatility forecasting.

Pipeline:

  1. Returns: r_t = (y_t - y_{t-1}) / y_{t-1}

  2. AR(p) on full returns gives mean return mu_t

  3. GARCH(p,q) on full returns gives sigma2_GARCH

  4. Ground-truth variance: sigma2_t = (r_t - mu_t)^2

  5. LSTM trained on sigma2_t sequences with GINN loss

Usage

GINN(
  data,
  mode = c("auto", "manual"),
  ar_max_lag = 5L,
  ar_lag = NULL,
  garch_p_grid = 1:3,
  garch_q_grid = 1:3,
  garch_mean_grid = c("constant", "ar", "arma"),
  garch_dist_grid = c("norm", "std", "ged", "snorm", "sstd", "sged"),
  garch_p = NULL,
  garch_q = NULL,
  garch_mean = NULL,
  garch_dist = NULL,
  lag,
  hidden_size = 32,
  num_layers = 1,
  lr = 0.001,
  dropout = 0.0,
  epochs = 500,
  patience = 30,
  tune_epochs = 50,
  batch_size = 16,
  lambda_list = seq(0.1, 0.9, 0.1),
  hidden_grid = c(8, 16, 24, 32, 40),
  layers_grid = c(1, 2),
  lr_grid = c(0.001, 0.003, 0.005),
  dropout_grid = c(0.0, 0.1, 0.2, 0.3),
  verbose = TRUE,
  seed = 42L
)

Arguments

data

Numeric data vector. ALL observations used for training. GINN internally computes returns and conditional variance.

mode

"auto" for automatic AR + GARCH + LSTM tuning. "manual" for user-supplied parameters.

ar_max_lag

Max AR lag to search (auto mode). Default 5.

ar_lag

Fixed AR lag (manual mode). NULL = auto.

garch_p_grid

ARCH order grid (auto mode). Default 1:3.

garch_q_grid

GARCH order grid (auto mode). Default 1:3.

garch_mean_grid

Mean model candidates for grid search: "constant" (intercept only), "ar" (AR(p) mean), "arma" (ARMA mean). Default c("constant", "ar", "arma").

garch_dist_grid

Error distribution candidates: "norm" (normal), "std" (Student-t), "ged" (GED), "snorm" (skew-normal), "sstd" (skew-t), "sged" (skew-GED). Default includes all six.

garch_p

Fixed ARCH order (manual mode).

garch_q

Fixed GARCH order (manual mode).

garch_mean

Fixed mean model (manual mode): "constant", "ar", or "arma".

garch_dist

Fixed distribution (manual mode).

lag

Variance sequence look-back length. COMPULSORY – no default. User must specify. Example: lag = 5 uses the last 5 variance values to predict the next one.

hidden_size

LSTM hidden units (manual mode). Default 32.

num_layers

Stacked LSTM layers (manual mode). Default 1.

lr

Learning rate (manual mode). Default 0.001.

dropout

Dropout rate (manual mode). Default 0.0.

epochs

Max training epochs. Default 500.

patience

Early-stopping patience. Default 30.

tune_epochs

Epochs per grid-search combo (auto mode). Default 50.

batch_size

Mini-batch size. Default 16.

lambda_list

Lambda values to test. Default seq(0.1, 0.9, 0.1). lambda = 0.0: Standard LSTM (GT variance only, no GARCH guidance). lambda = 0.5: Balanced GINN (equal GT and GARCH weight). lambda = 1.0: GINN-1 (GARCH variance only).

hidden_grid

Hidden-unit candidates (auto mode).

layers_grid

Layer-count candidates (auto mode).

lr_grid

Learning-rate candidates (auto mode).

dropout_grid

Dropout candidates (auto mode).

verbose

Print progress messages. Default TRUE.

seed

Random seed. Default 42.

Value

S3 object of class "GINN" with elements:

results

Named list per lambda. Each entry contains: tr_rmse, tr_mae, tr_r2, tr_preds, tr_actuals (all in VARIANCE units), val_rmse, loss_hist, time, lambda, model.

best_model

Complete best lambda summary: lambda, hyperparams, ar_lag, garch_config, metrics table, predictions (Actual_Var vs Predicted_Var data frame), loss_history, model object, train_time.

garch

GARCH config, sigma2_all (full in-sample variance), coefs.

ar

AR lag and fitted mean returns mu_all.

returns

Full return series r_t.

variance

Ground-truth variance sigma2_t (all observations).

best_hp

Best LSTM hyperparameters found.

meta

Settings: lag, lambda_list, mode, best_lam_nm, selection = "Train RMSE".

data_info

prices, N, N_ret, var_scaler, last_var_seq_sc, lag, mu_all.

Examples


prices <- cumprod(1 + rnorm(60, 0.001, 0.02)) * 100

# -- Fast manual mode (fixed GARCH order + small LSTM) -------
fit <- GINN(prices, mode = "manual", lag = 5,
            ar_lag = 1, garch_p = 1, garch_q = 1,
            garch_mean = "constant", garch_dist = "norm",
            hidden_size = 4, num_layers = 1, lr = 0.01,
            epochs = 20, patience = 5,
            lambda_list = c(0.5), verbose = FALSE)
print(fit)
summary(fit)

# -- Forecast 12 steps of conditional variance --------------
fc <- predictGINN(fit, h = 12)
print(fc)

# Volatility = sqrt(variance)
sqrt(fc$forecast_var)

# -- Diagnostic plots ---------------------------------------
plotGINN(fit, h = 12)      # actual + predicted + forecast
plotlambda(fit)             # Training RMSE by lambda

# -- Best model details -------------------------------------
fit$best_model
fit$best_model$lambda
fit$best_model$hyperparams
fit$best_model$metrics
fit$best_model$predictions   # Actual_Var vs Predicted_Var

# -- Manual mode with a couple of lambdas --------------------
fit2 <- GINN(
  prices,
  mode       = "manual",
  lag        = 5,
  ar_lag     = 1,
  garch_p    = 1,
  garch_q    = 1,
  garch_mean = "ar",
  garch_dist = "norm",
  hidden_size = 4,
  num_layers  = 1,
  lr          = 0.01,
  epochs      = 20,
  patience    = 5,
  dropout     = 0.0,
  lambda_list = c(0.3, 0.7),
  verbose     = FALSE
)
summary(fit2)

# -- Accuracy vs actual future prices ----------------------
new_prices <- tail(prices, 1) * cumprod(1 + rnorm(12, 0.001, 0.02))
AccuracyGINN(fit, actual = new_prices)

# Plot with actual future variance overlaid in green
plotGINN(fit, h = 12, actual = new_prices)

Plot GINN Volatility: Actual vs Predicted + Forecast

Description

Plots conditional variance (volatility) in one chart showing:

  1. Actual realized variance sigma2_t on full training data (dark line).

  2. GINN in-sample predicted variance (blue line).

  3. Out-of-sample variance forecast (red solid line).

If actual future prices are provided, their realized variance is computed and overlaid as a green line for direct comparison with the red forecast.

Usage

plotGINN(fit, h = 12, lambda = NULL, actual = NULL,
          newdata = NULL, n_hist = NULL)

Arguments

fit

A GINN object returned by GINN().

h

Forecast horizon (steps ahead in variance space). Default 12.

lambda

Lambda to use. NULL = best lambda.

actual

Optional numeric vector of actual future PRICES. Their variance is computed and overlaid in green.

newdata

Optional new prices to update starting sequence.

n_hist

Number of historical variance points to show. NULL = all.

Value

A ggplot2 object (invisibly).

Examples


prices <- cumprod(1 + rnorm(60, 0.001, 0.02)) * 100
fit    <- GINN(prices, mode = "manual", lag = 5,
                ar_lag = 1, garch_p = 1, garch_q = 1,
                garch_mean = "constant", garch_dist = "norm",
                hidden_size = 4, num_layers = 1, lr = 0.01,
                epochs = 20, patience = 5,
                lambda_list = c(0.5), verbose = FALSE)

# Plot variance forecast only
plotGINN(fit, h = 12)

# Show last 50 historical variance points only
plotGINN(fit, h = 12, n_hist = 50)

# With actual future prices overlaid (green = actual variance)
new_prices <- tail(prices, 1) * cumprod(1 + rnorm(12, 0.001, 0.02))
plotGINN(fit, h = 12, actual = new_prices)

Plot Training RMSE Across Lambda Values (GINN)

Description

Plot Training RMSE Across Lambda Values (GINN)

Usage

plotlambda(fit)

Arguments

fit

A GINN object.

Value

ggplot2 object (invisibly).

Forecast Conditional Variance h Steps Ahead from GINN

Description

Forecasts future conditional variance (volatility) h steps ahead using recursive one-step prediction on variance sequences. Returns single-point forecasts only.

Usage

predictGINN(fit, h = 1, lambda = NULL, newdata = NULL)

Arguments

fit

A GINN object returned by GINN().

h

Number of steps to forecast ahead. Default 1.

lambda

Lambda value to use. NULL = best (lowest Train RMSE). Use 0.0 for Standard LSTM (GT variance only, no GARCH guidance).

newdata

Optional numeric vector of new DATA that arrived after training. Their variance is computed and used to update the starting sequence before forecasting.

Value

Data frame with columns:

h

Forecast horizon step (1, 2, ..., h).

time_idx

Time index in return space (N_ret + 1, ..., N_ret + h).

forecast_var

Point forecast of conditional variance sigma2_t.

Examples


prices <- cumprod(1 + rnorm(60, 0.001, 0.02)) * 100
fit    <- GINN(prices, mode = "manual", lag = 5,
                ar_lag = 1, garch_p = 1, garch_q = 1,
                garch_mean = "constant", garch_dist = "norm",
                hidden_size = 4, num_layers = 1, lr = 0.01,
                epochs = 20, patience = 5,
                lambda_list = c(0.3, 0.7), verbose = FALSE)

# Forecast 12 steps of conditional variance
fc <- predictGINN(fit, h = 12)
print(fc)

# Volatility = sqrt(variance)
sqrt(fc$forecast_var)

# Specific lambda
fc3 <- predictGINN(fit, h = 6, lambda = 0.3)

# With new prices to update starting sequence
new_prices <- tail(prices, 1) * cumprod(1 + rnorm(5, 0.001, 0.02))
fc_new     <- predictGINN(fit, h = 12, newdata = new_prices)

mirror server hosted at Truenetwork, Russian Federation.