The thesis dataset contains only 75 dates per year and eight categories; conclusions should
not be generalized to large or multi-echelon supply chains.
Publication rights for the source sales data are unconfirmed, so reproducible public examples
use synthetic data rather than the original observations.
Weather can improve or degrade forecasts depending on category. The ablation is empirical and
cannot establish a causal effect.
Rule count grows exponentially with input count. The six-input weather model permits up to 64
rules even with only two membership functions per input.
The reference EFS training path depends on licensed MATLAB toolboxes. The independent Python
IT2 backend removes that runtime dependency but is not a claim of numerical equivalence to
MathWorks' interval type-2 implementation.
Both EFS backends can be substantially slower than simple statistical baselines in the full
three-seed optimization profile.
LSTM estimates are unstable on very small datasets; multiple seeds and uncertainty summaries
are required.
The current system retrains in batches. It does not yet evolve online, consume live streams,
optimize inventory decisions, or forecast multiple supply-chain echelons.
The Streamlit application is a local research interface, not a hardened multi-user production
service.