Member Avatar for GiorgosMaragkop

I am currently building a LSTM model to predict only the daily Open price of a stock.

Is there a difference in the prediction of the Opening price if I include other parallel series (High, Low, Close, technical indicators etc) using a Multiple Parallel Series Model compared to using a univariate LSTM only on Open prices?

In other words, a Multiple Parallel Series LSTM with N series, just runs N different univariate LSTMs?

Thank you in advance!

The code for Multiple Parallel Series LSTM is from this tutorial: https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/

  • 2 Contributors
  • 1 Reply
  • 312 Views
  • 1 Hour Discussion Span
  • Latest Post Latest Post by rproffitt

Dani AI

Generated

Short answer: feeding Open + High/Low/Close/indicators into a single multivariate LSTM is not the same as running N independent univariate LSTMs. A multivariate LSTM receives a 3‑D tensor (samples, timesteps, features) and can learn both temporal patterns and cross‑feature interactions. Whether that improves Open‑price forecasts depends entirely on whether the extra series contain predictive signal and on proper training/validation.

Architectures that are commonly compared:

  • single multivariate LSTM (one LSTM sees all features),
  • parallel/branched LSTMs (one LSTM per series, then concatenate outputs),
  • ensembles of independent univariate LSTMs (separate models combined at prediction time).

Minimal Keras conceptual examples (shapes only):

# single multivariate
X.shape == (n_samples, timesteps, n_features)
model.add(LSTM(50, input_shape=(timesteps, n_features)))
model.add(Dense(1))
# parallel branches (conceptual)
inA = Input(shape=(timesteps,1)); a = LSTM(32)(inA)
inB = Input(shape=(timesteps,1)); b = LSTM(32)(inB)
m = concatenate([a,b]); out = Dense(1)(m)

Practical tips and pitfalls: prefer modeling returns or scaled prices rather than raw levels; fit scalers only on training data; avoid any indicator that leaks future information; evaluate with time‑aware methods (walk‑forward or expanding‑window validation), not random CV; include a simple persistence baseline (last value) and report MAE/MSE plus directional hit‑rate. Apply regularization (dropout, early stopping), perform feature‑ablation or permutation importance to check whether extra series actually help, and consider dimensionality reduction if indicators are many.

A final note echoing : market moves are often driven by exogenous news and regime shifts that these models cannot foresee. The recommended approach is empirical: compare univariate baseline, single multivariate, branched, and ensemble approaches on out‑of‑sample data and prefer the simplest architecture that gives robust, repeatable improvement.

Member Avatar for rproffitt

Since you are predicting and can't take into account real world impacts like legal wins and losses along with market sentiment all these are fairly compared to junk science. I'd google a bit more and not put money down on such predictions. That said, there are those that hype and sell these stock market apps to help you pick the winners or if you are out to short a stock, the losers.

What has worked for me over the years are companies with a story. Microsoft, Cisco, Red Hat, Amazon, Apple were good bets I took. Today it's Tesla but your MPS LSTM won't get this one right as this one is the story of bears and competing industries. Old industries are going out of their way to maintain the status quo with FUD along with legal wranglings to stifle and slow the change.

If you are looking for this to predict the market, it's far too short of the mark.

Be a part of the DaniWeb community

We're a friendly, industry-focused community of developers, IT pros, digital marketers, and technology enthusiasts meeting, networking, learning, and sharing knowledge.