FinTime-Fuse beats every text-free Bitcoin baseline even when its news is shuffled: MSE is 4.023 against Time-LLM's 4.207 and PatchTST's 4.295. Shuffling even improves MAE and MAPE relative to correctly dated news (1.386 against 1.390, and 13.829% against 14.871%). Squared error favors alignment. Correctly dated news brings MSE down to 3.582; by our arithmetic, alignment supplies roughly 71% of the MSE improvement over Time-LLM.

Adding a FinBERT text path to Time-LLM

Time-LLM supplies the starting architecture. It divides a price series into patches, then maps them into a frozen language model's embedding space. Cross-attention connects each patch to a small collection of prototype vectors drawn from pretrained word embeddings. Those vectors anchor local shapes such as "short up". Mou, Xue, Chen, Takiguchi, Harada, Kamihigashi and Ariki retain this numeric path, using patch length 8 and stride 8 across a 96-day input window.

The addition is text. Frozen FinBERT reduces each day's text to a single 768-d vector, its [CLS] token. A linear projection maps that vector into LLaMA-7B's 4096-d space. The model concatenates text and price tokens and sends them through 8 frozen LLaMA layers. A linear head uses the last patch positions to produce the full 12-day forecast at once. Training updates only the projection, reprogramming and head layers.

The textual inputs differ by market. Gold has 1,352 daily observations from 2019 to 2023; FX has 2,347 observations covering USD against EUR, JPY and GBP from 2011 to 2019. Both use a templated macro-calendar feed. Bitcoin has 1,234 observations from 2018-01-03 to 2021-05-20 and uses ChatGPT-4 summaries. Those summaries draw on 3,495 article-level records retained from 5,827 raw timestamped articles. The chronological split is 70/10/20, and validation MSE determines the checkpoint.

The headline Bitcoin MSE is 3.582 versus Time-LLM's 4.207, a reduction of about 14.9%. Gold comes in at 0.041 against 0.054. FX gives Time-LLM a slight lead in the unrounded figures (0.0143700 versus 0.0144249). As the authors acknowledge, the model "does not uniformly dominate every metric in every setting".

These are forecasts of a normalized 12-step price path. The paper reports every metric "on the normalized scale used during model training, before inverse transformation." It reports no positions, returns or costs. The comparison of 3.582 with 4.207 therefore measures forecast error alone.

What does dating the news correctly buy?

Bitcoin's random-news variant records MSE of 4.023, between aligned news at 3.582 and Time-LLM at 4.207. On our arithmetic, that assigns roughly 71% of the squared-error improvement over Time-LLM to alignment. Absolute error gives a different answer: shuffled news achieves MAE of 1.386 against 1.390 for aligned news. Its MAPE is also lower, 13.829% versus 14.871%. That is the lowest MAPE anywhere in the Bitcoin comparison, including PatchTST's 14.234%.

The authors address this result directly. Aligned news retains the lower MSE and RMSE, they argue, "therefore, the full aligned-news setting remains the strongest configuration under the main MSE-based evaluation criterion." Choosing that criterion is fair. In our reading, an MSE difference of 0.441 beside an MAE difference of 0.004 suggests the advantage comes from large-error days. RMSE supports that reading: 2.006 with shuffled news against 1.893 with aligned news.

The test block contains roughly 250 Bitcoin days (20% of 1,234), with fewer 12-day forecast windows. The paper acknowledges a "unified single-run protocol rather than repeated multi-seed evaluation." The authors describe the gains as empirical evidence from a controlled experimental setting and defer the multi-seed stability check to future work. This is an honest framing. It still leaves the 0.441 MSE advantage from alignment untested against seed variation, while nearly identical MAE supplies no support from a second metric. One seed and roughly 250 test days offer thin grounds for an advantage visible only in squared error.

Gold gives a different result. Random news scores 0.053 and bert-base-uncased text scores 0.055, both worse than the no-text result of 0.049. Alignment helps only when paired with the FinBERT encoder.

The Bitcoin no-text ablation proves too much

Without text, Bitcoin MSE jumps to 27.927 and MAPE reaches 38.671%. The paper calls this a 679.73% increase, suggesting "that textual information provides complementary signals beyond historical numerical patterns, especially for the Bitcoin dataset". We draw a different inference. Time-LLM is essentially the same architecture with the text path removed, yet its MSE is 4.207. A text-free variant performing 6.6 times worse than Time-LLM looks to us like a degraded configuration. We would not use it to price the news. Although the paper says it holds "the remaining architecture and training configuration unchanged", we did not find an account of what occupies the text positions after text is removed.

The comparison against Time-LLM prices the text at 14.9%.

Reproducing it leaves the scaler unresolved

The paper pins down several choices: patch size, the 96/12 window and horizon, a learning rate of 5e-4, batch size 1, patience 10, and the 8 LLaMA layers. Other choices require judgment.

Algorithm 1 calls LLaMA three times. The first call uses text alone; another uses the prompt plus patches. It then concatenates the last-layer outputs from both calls and feeds them back as input embeddings. Using final hidden states as fresh inputs is unusual. A reproducer has to decide whether to follow that instruction literally.

Normalization requires another decision. We did not find the scaler's identity or whether its statistics are fitted only on training data. The test block covers the final 20% of the sample ending on 2021-05-20. Our arithmetic puts its start around September 2020. If the scaler uses the training block alone, Bitcoin RMSE of 1.893 in normalized units suggests test prices were well outside the training range. Hyperparameters vary across markets too: d_model is 64 for Bitcoin, compared with 768 for gold and FX.

The text warrants care as well. On Bitcoin's 186 days without news, the input is the literal string "no news". A keyword audit detected realized-price wording in 492 of the 3,495 source records and in 14 summaries actually read by the model. The authors describe these as heuristic matches and say the expressions "were not systematically removed". Because the text stays within the input window, these passages repeat price history already available to the model. The authors also state that public data sources do not consistently supply exact intraday release times. They accordingly frame the exercise as multi-day forecasting.

Our Bitcoin run uses a numeric-only baseline

We cannot train a LLaMA-7B fusion model. Our run takes a numeric-only baseline from the paper's comparison set and uses Bitcoin alone; it does not test FinTime-Fuse. We have no FX data. We are also not substituting an ETF for the paper's gold series. Before computing any return, we fix the trading rule: the sign of the 12-day forecast relative to the last close determines the position. We then measure returns net of costs.

An aligned-text model that trades better than a shuffled-text model would change our view. Our numeric-only run cannot answer that comparison.