Should you evaluate Kronos?
Choose Kronos when financial candles are the object of your research and you can verify the evaluation dates against the released checkpoint's training history. Start with a baseline comparison before investing in fine-tuning. The supplied evidence establishes a documented forecasting interface and an unsuccessful sandbox test command; it does not establish profitable trading or superior forecasts on your data.
The README describes the included Qlib pipeline as a simplified demonstration rather than a production-ready quantitative trading system. Its production considerations include portfolio construction, risk controls, transaction costs, slippage and market impact. Those are additional evaluation and implementation tasks.
What our sandbox established
In our October 4, 2026 sandbox run at commit 67b630e, installation succeeded in 63 seconds. The attached harness records 80 dependency packages and a 5,519 MB footprint. The build step returned zero in 1 second, but its empty output tail provides no evidence of a completed forecasting workload.
The test command exited 1 after 313 seconds. Its stored parser summary is 2 passed, 2 failed, 1 error and 0 skipped, with total 5. These counts have not been reconciled against complete logs to establish exhaustive unique test coverage.
The raw footer separately identifies timeouts in test_kronos_predictor_mse[512-0.008979] and test_kronos_predictor_mse[256-0.003741], both in tests/test_kronos_regression.py. Each exceeded pytest's 120-second limit. It also names ERROR finetune/qlib_test.py, without enough detail to assign a cause. The raw footer reports 308.10 seconds; the attached harness records 313 seconds for the command. These are different recorded timings, and neither replaces the other.
Our harness stores command success, exit codes, elapsed time, timeout status, output tails and parsed test counts. The MrKeyoor testing-methodology page explains that reporting approach and its limits. It is provenance context, not evidence that this older run used today's procedure. This correction uses the existing record and does not report a new run.
Pip-audit recorded 0 advisories. The record does not include scannedPackages, so package scan coverage is unknown. A zero advisory count does not establish security. Forecast quality, returns after costs, Windows behavior, sustained service operation and multi-GPU fine-tuning were not established by our run.
Confirm checkpoint provenance before choosing test dates
The supplied README describes training across more than 45 exchanges but does not specify the corpus's temporal range. Issue 265 asks for training dates, US equity coverage and a clean evaluation window.
A September comment in that issue says the paper describes pretraining through June 2024, then asks whether the released Hugging Face weights share that cutoff. That is a commenter's account and an unanswered provenance question in the supplied extract. It does not confirm the cutoff of the checkpoint you download. Another comment alleges leakage; the supplied evidence does not establish that allegation as a finding.
Record the exact model and tokenizer revisions, then seek confirmation of their latest training dates and asset coverage. Until that is resolved, label a historical comparison's pretraining overlap as unknown rather than calling its dates unseen.
Match the documented interface to your data
Kronos requires historical open, high, low and close columns, historical timestamps and future timestamps. Volume and amount are optional. The predictor returns forecast values in a pandas DataFrame. Sampling settings include temperature, top-p and sample count; record them with the checkpoint revisions so another researcher can reproduce the configuration.
The model table lists mini with a 2,048-step context and small and base with 512 steps. The predictor truncates longer inputs for small and base. Batch prediction requires equal history lengths and equal forecast lengths across series. The listed large checkpoint has no available weights in the supplied table.
Fine-tuning is a separate setup commitment. The example uses daily Chinese A-share data prepared through Qlib, configurable dataset and checkpoint paths, and successive tokenizer and predictor training stages. Other data formats require adapting the example's loading and preprocessing logic. The README also warns that generated comments in the fine-tuning directory may contain inaccuracies; check the implementation when a comment affects an experimental decision.
Choose comparison tools by the question
Chronos-2 supports zero-shot univariate, multivariate and covariate-informed forecasting. Its repository makes it a candidate when your experiment needs a general forecasting model or external variables. That documented scope does not establish better performance on financial candles.
NeuralForecast provides a collection of neural forecasting models with fit and predict interfaces. Use its repository when you want to train several neural baselines under a common interface. It can complement a Kronos evaluation rather than replace the pretrained checkpoint. Neither alternative supplies the missing provenance for Kronos or validates a trading strategy.
For the next experiment, fix the evaluation dates and inputs, include a last-value baseline, and compare forecast errors before adding a trading rule. If you evaluate a strategy, specify costs and execution assumptions separately. Our sandbox results cannot answer either performance question.
How to read the editorial ratings
These ratings describe the supplied evidence, not benchmark performance. The rubric is: 1 means major barriers, 2 means substantial unresolved barriers, 3 means usable with material caveats, 4 means strong documented support with a specific remaining gap, and 5 means comprehensive evidence for the dimension. Unknown dimensions remain unassessed.
Setup ease receives 2 because installation succeeded but the footprint was substantial and the test command failed. Documentation receives 4 for its inference, batch, configuration and fine-tuning instructions, with released-checkpoint provenance still unresolved. Maturity receives 2 because our recorded checkout had no detected CI workflows and our test command was unsuccessful; this is not a judgment about forecast accuracy. Community is unassessed because the supplied primary extracts do not establish participation trends, maintainer responsiveness or current adoption.