Assessing Multi-Horizon LSTM Forecasts for Commodity Hedging

Summary

The document describes a forecasting setup for daily commodity prices using an LSTM network and predictor variables drawn from current and lagged observations. The author asks whether forecast accuracy can determine if the model is suitable for managing hedging operations, particularly when deciding whether to enter a forward contract.

The question reports approximately 92% accuracy at a one-day horizon and 39% at a one-month horizon, highlighting that performance falls as the forecast horizon grows. These figures are not accompanied by a definition of accuracy, benchmark comparisons, out-of-sample evaluation, or details about the hedging objective. No answer establishes whether the forecasts support a hedge decision. The document points toward the need to assess forecasts in the context of the actual risk reduction and costs of hedging, rather than relying on an accuracy percentage alone.

Key ideas

• The forecasting setup uses an LSTM to predict commodity prices from present and lagged predictor data.
• Reported accuracy declines from the one-day horizon to the one-month horizon.
• The author asks whether one-month forecast performance can guide forward contract hedging decisions.
• The document does not define accuracy or provide benchmarks, hedge outcomes, or an answer.

Tags

# Forecasting accuracy in one month and hedging


# Forecasting accuracy in one month and hedging












I am working on predicting the daily data of a financial time series $[Y(t+1),...Y(t+j)]$ =$f(X_1(t),...X_1(t-i),.....,X_n(t),...X_n(t-i))$ where $Y$ is a commodity price $X_i$ are predictor variables and I used LSTM (Long Short-Term Memory networks) in order to do that...

How to judge if this predictor is adequate to the task of managing hedging operations?

Specifically I am wondering if an accuracy of one month forecast of 39% is a good one in order to take a decision to buy a forward contract in order to do hedging?

PS : my one step accuracy (one day) is about 92% but as you know when the forecast horizon is bigger the accuracy goes down!

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.