Razzaq and Majeed's tables offer no reason to size a trade on Twitter or Telegram sentiment. Their clearest comparison runs against their conclusion: on the Telegram sample, ARIMA has lower RMSE than LSTM and Bi-LSTM for Bitcoin, Ether and Binance Coin.

From messages to forecasts

The premise is straightforward. If social-media mood moves with crypto prices, a daily sentiment index should help predict daily returns. The authors use Yahoo Finance prices for the three coins from July 19, 2023 to July 19, 2024, calculating simple daily returns. They collect tweets through the Twitter API using hashtag and keyword filters. Telegram messages come through its API and scraping of public and private channels. VADER, a lexicon-based scorer tuned for slang and emoji, assigns sentiment scores to both sources; the authors aggregate these daily and merge them with returns by date.

They then conduct four analyses. Pearson correlations and covariances come first, followed by ARIMA, whose order is selected by AIC, and SARIMA. LSTM and Bi-LSTM networks use 70% of the data for training and the remaining 30% for testing, with grid search setting the hyperparameters. Historical-quantile VaR at 90, 95 and 99% completes the analysis. Each analysis is repeated on the merged Twitter sample and the Telegram sample.

The authors argue that Twitter supplies immediate, short-term information, whereas Telegram captures broader, noisier trends. They say neural networks outperform the time-series models and that combining the platforms "can improve predictive accuracy." They acknowledge that ARIMA and SARIMA "were better in individual circumstances but overall failed to be successful." The claim hangs on those words "but overall".

How much data reaches the models?

Four sample sizes appear in the paper. They cannot all refer to the same dataset. The stated price period spans a year, yet the coin data are brought into "the standard scope of 230 days." Telegram covers December 22, 2022 to July 20, 2024 and contains 8,537 records. Of the combined coin records, 8,513 were matched to it. The paper then says: "After filtering, the final datasets for Bitcoin, Ether, and Binance Coin each comprised 16 data points, while the sentiment data encompassed 18 data points."

If the forecasting models use that 16-point set, the 70/30 split leaves about five test days and roughly eleven training observations for a grid-searched LSTM. An RMSE ranking based on five test observations has too little data to be statistically meaningful. We did not find a Diebold-Mariano test in the paper, nor a zero-return or random-walk forecast for the models to beat.

The platform comparison changes the return sample

The correlation tables compare coins with other coins. Neither includes sentiment. BTC-ETH correlation in the Twitter sample is 0.90935, with covariance of 209.4299. For the same pair in the Telegram sample, the figures are 0.7339 and 4.531258.

Sentiment has no place in a correlation calculated between two price-return series. These are the same coins over the same stated year, yet one covariance is about 46 times larger. The return samples must differ, or their scaling must.

The VaR results reveal the same problem. Bitcoin's 99% VaR is -11.17% for Twitter and -5.79% for Telegram; Ether's is -17.52% and -5.83%, respectively. As described, ARIMA forecasts historical returns, and we found no sentiment regressor attached to it. Even so, Bitcoin's ARIMA RMSE changes from 12.7371 on Twitter to 0.5228 on Telegram, a gap of about 24 times. Whatever else the platform split does, it changes the returns fed to the models. Each Twitter-versus-Telegram comparison therefore confounds the information source with the sample.

The promised Pearson correlation between sentiment and returns never appears in the results. There is no lead-lag test or regression either. Instead, the sentiment claim depends on dotted plots in Figures 2 to 7 and the assertion that positive spikes "usually coincided" with higher returns. The text also misreads Table 3: it gives ARIMA credit for Bitcoin's 17.6074 RMSE and 185.53% MAPE, although the table attributes those figures to SARIMA.

Which metric decides the winner?

The Telegram RMSE results favor ARIMA throughout:

MAPE reverses the result for every coin. Binance Coin's LSTM records 0.52%, compared with ARIMA's 13.01%. Across all twelve Telegram results, RMSE ranges only from 0.5024 to 0.6393. MAPE spans 0.52% to 36.46%.

Standard MAPE divides each error by its realized value. The paper describes a different calculation, total error divided by total absolute value, and the printed formula is garbled. Which calculation generated the reported figures remains unclear. Daily returns repeatedly cross zero. With standard MAPE, a return of a few basis points can swell a model's error regardless of its performance on the other days. In a test covering about five days, that observation can determine the ranking.

RMSE avoids that denominator problem. By this measure, ARIMA wins every coin in the Telegram sample.

The authors' "overall" judgement does fit the Twitter sample, where the networks beat the time-series models on both metrics. Ether's Bi-LSTM records 1.5492 and 33.82%, versus ARIMA's 14.8080 and 185.25%. The samples thus give opposite RMSE verdicts. If the models used the 16-point set, neither sample has enough test observations to resolve the question. An "overall" judgement requires one of them to do so.

The combined signal remains untested

The conclusion says that joining Twitter's immediacy to Telegram's breadth can improve forecasting accuracy. Yet every results table treats the platforms separately. We did not find a model using both sentiment series as inputs. Nor did we find a horizon comparison supporting the distinction between short-term Twitter information and long-term Telegram information.

We could not backtest this.

A replication requires historical Twitter and Telegram text for 2023 and 2024, and we have neither. Substituting crypto news sentiment would test whether editorial tone predicts returns. That tone comes from a different crowd working on a different clock, so such a test would not replicate this paper.

One result would change my mind: a model using both sentiment series as regressors, fitted on the full year, that beats a zero-return forecast on out-of-sample RMSE and reports a test statistic. Until then, these tables support only the conclusion that ARIMA won the Telegram sample on RMSE.