ferrifuzz.fts
Chen (1996) fuzzy time series: partition the value range into fuzzy sets, learn which set follows which, forecast by averaging. First order (FOFTS) and high order (HOFTS). The method is explained step by step in Fuzzy time series; the enrollment forecasting tutorial is a full worked example.
from ferrifuzz.fts import FuzzyTimeSeries
FuzzyTimeSeries
FuzzyTimeSeries(
n_partitions=7,
order=1,
mf_type="triangle",
partitioner="equal_width",
margin=0.1,
universe=None,
)
| Parameter | Default | Description |
|---|---|---|
n_partitions | 7 | number of fuzzy sets over the universe; must be ≥ 2 |
order | 1 | how many past steps form a rule antecedent. 1 = first order, ≥ 2 = high order; must be ≥ 1 |
mf_type | "triangle" | partition shape: "triangle" or "gaussian"; anything else raises ValueError |
partitioner | "equal_width" | partitioning strategy; "equal_width" is the only one implemented |
margin | 0.1 | when universe is not given: fraction of the data range added below the minimum and above the maximum |
universe | None | explicit (lo, hi) universe of discourse; overrides the data range and margin |
Invalid arguments (n_partitions < 2, order < 1, an unknown mf_type or
partitioner, a universe with lo ≥ hi) raise ValueError.
Because every value is assigned to its highest‑membership partition and both shapes peak
at the same centers, mf_type does not change the forecasts; it only changes how the
partitions look when plotted.
Partitions are equal width: centers with
. Triangles reach from one neighbour's center to the
next; Gaussians use . Labels are A1 … An, low to high.
fit(series)
Learns the universe, the partitions and the fuzzy logical relationship groups (FLRGs).
Returns self, so it chains. The series must be longer than order; a NaN or infinite
value in the training series raises ValueError (fill or drop missing values first).
DATA = [13055, 13563, 13867, 14696, 15460, 15311, 15603, 15861, 16807, 16919, 16388,
15433, 15497, 15145, 15163, 15984, 16859, 18150, 18970, 19328, 19337, 18876]
model = FuzzyTimeSeries(n_partitions=7, order=1).fit(DATA)
print(model.summary())
# FuzzyTimeSeries(n_partitions=7, order=1, mf_type='triangle', fitted, 5 FLRGs)
# Universe : [12426.8000, 19965.2000]
# Partitions : 7 (triangle)
# FLRGs found : 5
forecast(series): one step ahead
model.forecast(series) -> list[float]
For each , forecasts series[t] from series[t-order:t] only (no
look‑ahead). The result has len(series) - order values, aligned with series[order:]:
y_pred = model.forecast(DATA)
y_true = DATA[model.order:]
predict(history, steps=1): multi step
model.predict(history, steps=1) -> list[float]
Forecasts steps values past the end of history (which needs at least order
values). Each forecast is fuzzified and fed back in as input for the next step, so a
first‑order model quickly settles on one value (here partition only ever leads to
itself).
print(model.predict(DATA[-3:], steps=3)) # [18708.8, 18708.8, 18708.8]
evaluate(series)
model.evaluate(series) -> dict[str, float]
Shortcut for score(series[order:], model.forecast(series)): RMSE, MAE and MAPE of the
one‑step‑ahead forecasts on the aligned window.
print(model.evaluate(DATA)) # {'rmse': 535.81..., 'mae': 467.83..., 'mape': 2.87...}
To score on held‑out data, prepend the last order training values so the first test
point gets a forecast too: model.evaluate(train[-model.order:] + test).
Values outside the universe
A value outside the fitted universe (for example a test value larger than anything seen
in training) is mapped to the nearest partition: below the universe, above
it. Chen's model can therefore never forecast beyond the outermost partition centers.
If you know the plausible range in advance, set it with universe=(lo, hi) or widen
margin.
Inspection
| Member | Description |
|---|---|
get_partitions() | list[Partition], low to high |
get_flrgs() | list[FLRG] in order of first appearance |
summary() | multi‑line description (universe, partitions, number of FLRGs) |
repr(model) | one line, with fitted, k FLRGs or not fitted |
n_partitions, order, mf_type, partitioner, margin, universe | the constructor arguments |
forecast, predict, evaluate, get_partitions, get_flrgs and summary raise
RuntimeError if called before fit.
Partition
One fuzzy set of the universe.
| Member | Description |
|---|---|
label | "A1", "A2", … |
center | the partition's center (its MF peak) |
mf | the underlying Type1MF |
membership(x) | degree of x in this partition |
FLRG
A fuzzy logical relationship group: an antecedent window and every partition that followed it in the training data.
| Member | Description |
|---|---|
antecedent | tuple of partition indices (0‑based), length order |
consequents | list of partition indices that followed it (no duplicates) |
for g in model.get_flrgs():
print(g)
# FLRG(A2 -> A2, A3)
# FLRG(A3 -> A3, A4)
# FLRG(A4 -> A4, A5, A3)
# FLRG(A5 -> A4, A6)
# FLRG(A6 -> A6)
The forecast for a window is the mean of its consequents' centers. If a window never occurred in training, the forecast is the center of the window's last partition.