Advancements in machine learning are reshaping time series forecasting, with the M competitions serving as crucial benchmarks. The M4 and M5 competitions highlighted the strengths of methods like ES-RNN and LightGBM, respectively. This study compares the M4 competition winner (ES-RNN) and a LightGBM forecasting approach inspired by its prevalence among top M5 solutions. We evaluate their performance on a challenging subset of 2724 daily frequency time series from the M4 dataset. Using multiple error metrics (sMAPE, MAE, MAPE, RMSE), our results indicate that LightGBM outperforms ES-RNN in a slight majority (approx. 54–55%) of the series. However, a key finding is that LightGBM’s superior performance is conditional and strongly linked to a discernible trend in the time series. LightGBM demonstrates a significant advantage and higher forecast correlation for series with clear upward or downward trends. Conversely, the hybrid ES-RNN model exhibits greater robustness for series lacking strong trends or displaying higher variability. We introduce a quantitative trend measure and employ non-parametric statistical tests to statistically validate the significant impact of the trend on the correlation between LightGBM forecasts and actual values. We conclude that while LightGBM is highly effective, particularly for trended data common in domains like retail sales, its broader application requires careful consideration of the underlying data characteristics. Understanding the time series trend is crucial for practitioners selecting between these powerful forecasting methods.