MiniRocket: Deterministic Time-Series Transform

• MiniRocket is a time-series classification transform that uses a fixed, engineered kernel bank with two-valued weights, reducing randomness in feature extraction.
• It replaces the traditional global max pooling with a sole Proportion of Positive Values (PPV) statistic, significantly lowering computational cost while maintaining competitive accuracy.
• MiniRocket achieves up to 75× faster processing on large datasets, making it a strong baseline for scalable and efficient time-series analysis.

MiniRocket, short for Minimally Random Convolutional Kernel Transform, is a time-series classification transform in the ROCKET family that replaces fully random convolutional kernels with an almost deterministic, highly structured kernel bank, retains Proportion of Positive Values (PPV) as the sole pooling statistic, and trains a simple linear classifier on the resulting feature vector. In its original formulation, MiniRocket is up to 75 times faster than ROCKET on larger datasets while maintaining essentially the same accuracy, and it can train and test a classifier on all 109 datasets from the UCR archive in less than 10 minutes (Dempster et al., 2020). Standard implementations produce about 9,996 scalar features per series and have become a common baseline for fast time-series classification (Dempster et al., 2020, Lo et al., 2024).

1. Position within the ROCKET family

ROCKET introduced the now-standard pipeline of applying many 1D convolutional kernels to a time series, extracting simple global statistics from each activation map, and fitting a linear classifier on the transformed features. MiniRocket is a reformulation of that pipeline: it preserves the emphasis on dilation, PPV pooling, and linear classification, but removes most of the randomness in kernel generation and feature extraction, thereby shifting the method from a large random-feature transform toward an almost deterministic transform (Dempster et al., 2020, Lo et al., 2024).

The canonical contrast is that ROCKET uses many random kernels with random length, weights, bias, dilation, and padding, whereas MiniRocket fixes the kernel length to 9, uses a fixed family of kernels with two-valued weights, derives biases from convolution outputs rather than from a random uniform distribution, and discards global max pooling in favor of PPV alone (Dempster et al., 2020, Lo et al., 2024). Later papers summarize the weight restriction as values in $\{-1,2\}$, and some describe equivalent sign-flipped forms such as $\{-2,1\}$ (Lo et al., 2024, Hao et al., 3 Nov 2025).

Aspect	ROCKET	MiniRocket
Kernel construction	Fully random kernels	Fixed, engineered kernel set
Pooling	PPV + global max pooling	PPV only
Typical feature count	$20{,}000$ for $10{,}000$ kernels	About $9{,}996$ features

MiniRocket later became the basis for several descendants and reinterpretations. MultiRocket expands the set of pooling operators and input representations; SelF-Rocket adds wrapper-based selection over representations and pooling operators; HDC-MiniROCKET reinterprets MiniRocket in hyperdimensional-computing terms and augments it with explicit time encoding (Lo et al., 2024, Schlegel et al., 2022).

2. Kernel parameterization and feature map

For a univariate time series $\mathbf{X} = \{x_1,x_2,\ldots,x_T\}$, MiniRocket applies dilated 1D convolutions and converts each resulting activation map into a scalar PPV feature. A standard convolutional response can be written as

$(X * k)_t = \sum_{i=0}^{L-1} X_{t+id}\,w_i,$

and the PPV statistic for an activation map $Z=\{z_1,\dots,z_n\}$ is

$PPV(Z)=\frac{1}{n}\sum_{i=1}^{n}[z_i>0].$

In MiniRocket, the feature for kernel $k$ with associated bias $\{-2,1\}$0 is typically written as

$\{-2,1\}$1

and the standard feature vector contains approximately $\{-2,1\}$2 such features (Lo et al., 2024).

The original MiniRocket paper fixes the kernel length to 9 and uses a subset of 84 kernels drawn from the family of two-valued length-9 kernels. In the original construction, each kernel has weights $\{-2,1\}$3 and $\{-2,1\}$4, with exactly three positions assigned $\{-2,1\}$5 and the remaining six assigned $\{-2,1\}$6, yielding $\{-2,1\}$7 kernels from the relevant subset of the $\{-2,1\}$8 possible two-valued patterns (Dempster et al., 2020). Later summaries describe the same design more generally as a fixed set of 84 kernels with values in $\{-2,1\}$9, expanded by dilations into about $20{,}000$0 kernel/dilation combinations (Lo et al., 2024).

Bias generation is central to the method’s “minimal randomness.” Instead of drawing biases from $20{,}000$1, MiniRocket samples bias values from quantiles of convolution outputs. In the default version, for each kernel/dilation combination, the bias quantiles are taken from the convolution output on one randomly chosen training series; in the fully deterministic variant, the quantiles are computed from convolution outputs over the entire training set (Dempster et al., 2020). This detail means that MiniRocket is not strictly randomness-free in its default form, even though the kernel bank itself is fixed (Dempster et al., 2020).

The same transform is used beyond univariate classification. In prognostic classification on NASA’s CMAPSS data, the input is modeled as a multivariate time series

$20{,}000$2

and MiniRocket yields one PPV feature per kernel for each multivariate series (Wu et al., 2022). The method is also used on multivariate event-log-derived sequences for predictive maintenance and on raw or precomputed multichannel sensor data in human activity recognition (Vargas et al., 2023, Alagoz, 2024).

3. Determinism, computational structure, and efficiency

MiniRocket retains the same broad asymptotic transform complexity as ROCKET,

$20{,}000$3

but drastically reduces constant factors through kernel regularization, PPV-only features, and shared computation across kernels (Dempster et al., 2020). The original implementation exploits the two-valued kernel structure by precomputing $20{,}000$4 and $20{,}000$5, constructing a shared convolution component for the $20{,}000$6-weighted positions, and adding the contributions of the three $20{,}000$7-weighted positions for each kernel; the effect is that most work is shared across the 84 kernels for a fixed dilation (Dempster et al., 2020).

The reported speed gains are large. On the 109 UCR datasets, ROCKET required about 2 hours of total compute time on a single CPU core, including about 1h 55m for the transform, whereas MiniRocket required about 8 minutes total, including about 2.5 minutes for the transform (Dempster et al., 2020). On larger datasets such as MosquitoSound, FruitFlies, and InsectSound, MiniRocket was reported to be 43–75× faster than ROCKET (Dempster et al., 2020). The original paper therefore characterizes MiniRocket as significantly faster than any other method of comparable accuracy, and significantly more accurate than methods of roughly similar computational expense (Dempster et al., 2020).

The “almost deterministic” formulation also has a fully deterministic variant. Replacing per-kernel bias sampling from a random training series with bias estimation from the full training set yields a fully deterministic transform with empirically essentially identical accuracy, but at higher computational and memory cost (Dempster et al., 2020). In practice, the default MiniRocket configuration remains the usual choice.

Downstream classification is deliberately simple. The original MiniRocket work uses ridge regression on smaller datasets and logistic regression with Adam on larger datasets (Dempster et al., 2020). Later implementations and extensions commonly use RidgeClassifier as the default linear model, preserving the separation between a fixed convolutional transform and a lightweight classifier (Lo et al., 2024).

4. Accuracy profile and benchmark behavior

On the 109 UCR datasets, MiniRocket is reported as marginally more accurate on average than ROCKET, and only slightly less accurate than methods such as TS-CHIEF and HIVE-COTE/TDE, with the differences to those stronger ensembles reported as not statistically significant under the paper’s pairwise tests (Dempster et al., 2020). The core result is therefore not merely faster inference, but a favorable accuracy–cost trade-off (Dempster et al., 2020).

Later work clarified that MiniRocket’s strong average performance does not make PPV-only pooling universally optimal. A systematic study of 112 UCR datasets constructed 15 MiniRocket variants by combining five pooling operators with three input-representation sets and found that PPV_MIX had the best average mean rank, yet it was not best on most individual datasets and was outperformed by another pooling/operator combination in about 78.57% of datasets (Lo et al., 2024). This result motivated SelF-Rocket, which keeps MiniRocket’s kernel design but dynamically selects the best representation and pooling operator during training; on the same benchmark, SelF-Rocket ranked 2nd overall, below HIVE-COTE v2.0, and outperformed MiniRocket by about 1.23 percentage points in mean accuracy difference (Lo et al., 2024).

A different line of work exposed a structural limitation of MiniRocket’s global PPV pooling. HDC-MiniROCKET showed that MiniRocket can suffer catastrophic failures when class identity depends primarily on global temporal position. On a synthetic two-class dataset of noisy signals containing a single sharp peak, with class determined only by whether the peak lay in the first or second half of the series, standard MiniROCKET reached about 65.0% accuracy, while HDC-MiniROCKET with explicit time encoding reached 97.0%; on a harder subset, MiniROCKET dropped to about 56.8%, while HDC-MiniROCKET reached 94.1% (Schlegel et al., 2022). On 128 UCR datasets, HDC-MiniROCKET improved over MiniROCKET on 81 datasets when the time-encoding scale parameter $20{,}000$8 was chosen by oracle search, with an average gain of about 3.1% on those improved datasets (Schlegel et al., 2022).

These results collectively suggest a stable core conclusion. MiniRocket is a very strong default transform for general time-series classification, but its PPV-only, globally pooled representation can be suboptimal when either temporal localization or representation choice is critical (Schlegel et al., 2022, Lo et al., 2024).

5. Applications across domains

MiniRocket has been used well beyond the classical UCR benchmark. In prognostic health management on NASA’s CMAPSS turbofan dataset, MiniRocket was applied to health-status classification from multi-sensor time series. On FD001 with 500 kernels, MiniROCKET achieved 47.56% accuracy, 38.17% macro-F1, and 5 s computation time, compared with 46.26%, 41.13%, and 23 s for ROCKET, and much larger runtimes for stacked LSTM and GRU baselines. In a second experiment using 5000 kernels on FD001–FD004, MiniROCKET remained much faster than the deep baselines, with a reported computation time of 25 s, versus 111 s for ROCKET and about 460–526 s for the deep models (Wu et al., 2022). The same study also showed that MiniRocket/ROCKET features can be combined with SVM and LDA, and on the difficult FD004 subset, SVM and LDA on the random-kernel feature space reached 96.25% / 87.98% and 95.34% / 90.70% in accuracy/F1, respectively (Wu et al., 2022).

In predictive maintenance from ATM event logs, MiniROCKET was used after converting logs into multivariate time series of cumulative command–response counts in 10-minute windows. With a Ridge classifier, MiniROCKET achieved average test Accuracy $20{,}000$9, Balanced Accuracy $10{,}000$0, F1 $10{,}000$1, AUC $10{,}000$2, MS $10{,}000$3, and 23.3 s training time per fold (Vargas et al., 2023). The same study reported that MiniROCKET significantly outperformed InceptionTime on Balanced Accuracy, F1, AUC, and MS, while remaining statistically similar to ROCKET and much faster (Vargas et al., 2023).

In smartphone-based human activity recognition on the UCI HAR dataset, MiniRocket was evaluated both on the standard 561-dimensional handcrafted feature vectors and directly on raw single-channel sensor signals. On the precomputed-feature setup, MiniRocket achieved Accuracy $10{,}000$4, F1 $10{,}000$5, AUC $10{,}000$6, with a training time of 80.1 seconds; in raw single-channel mode, the best result came from the Y-axis total acceleration signal, with Accuracy $10{,}000$7, F1 $10{,}000$8, and AUC $10{,}000$9 (Alagoz, 2024).

MiniRocket has also been used successfully in EEG motor-imagery classification. On the PhysioNet MI-EEG dataset, a MiniRocket-plus-ridge-classifier pipeline achieved mean accuracy 98.63%, compared with 98.06% for a CNN–LSTM baseline, and the study explicitly states that classification via MiniRocket’s features achieved higher performance than the best deep learning models at lower computational cost (Hwaidi et al., 22 Aug 2025).

In hyperspectral spectral classification, MiniROCKET was used as a fixed feature extractor on per-pixel spectra. A 2025 study reported that MiniROCKET outperforms 1D-Justo-LiuNet in limited data scenarios and is mostly on par with it in the general case, with the strongest advantage below roughly 100,000 samples per class; on the HYPSO-1 dataset with full training data, MiniROCKET achieved OA 74.79, compared with 72.90 for 1D-Justo-LiuNet (Theisen et al., 17 Sep 2025).

6. Limitations, misconceptions, and ongoing directions

A common misconception is that “minimally random” means “fully deterministic.” In the original method, the kernel family, dilations, padding schedule, and feature type are fixed, but the default algorithm still samples bias quantiles from the convolution output of a randomly chosen training series for each kernel/dilation pair; full determinism requires a more expensive bias-estimation procedure over the full training set (Dempster et al., 2020).

A second misconception is that PPV-only pooling is universally sufficient. The evidence is mixed. The original paper found that PPV alone matched or slightly exceeded PPV plus max pooling, which justified dropping global max pooling (Dempster et al., 2020). However, subsequent work showed that pooling operators and input representations matter, and that MiniRocket’s global PPV can suppress information about where patterns occur in the series, producing failures on tasks in which global temporal position is decisive (Lo et al., 2024, Schlegel et al., 2022).

Classifier choice is another important practical issue. MiniRocket is often presented as a transform plus ridge classifier, but the PHM literature showed that the same feature space can benefit from SVM or LDA, particularly in very high-dimensional settings and on difficult subsets such as FD004 (Wu et al., 2022). This suggests that MiniRocket itself should be separated conceptually from any single downstream linear model.

Interpretability remains limited. The feature space is high-dimensional, individual PPV features act as generic pattern detectors, and application studies note the absence of explicit feature selection or dimensionality reduction in standard usage (Wu et al., 2022). In the hyperspectral setting, the fixed bias-fitting procedure was also identified as a possible source of suboptimal generalization because it relies only on the first batch for bias fitting in the reported implementation (Theisen et al., 17 Sep 2025).

Several active directions therefore extend rather than replace MiniRocket. SelF-Rocket introduces dataset-specific selection over pooling operators and input representations (Lo et al., 2024). HDC-MiniROCKET adds explicit time encoding while preserving MiniRocket’s convolutional front end (Schlegel et al., 2022). Other papers note the possibility of richer input representations, filter-based feature selection, and kernel-pruning approaches such as S-ROCKET, POCKET, and Detach-ROCKET (Lo et al., 2024). The general pattern is that MiniRocket remains the canonical fast baseline, while later work modifies its pooling, representation, temporal encoding, or feature-selection layers to address the cases in which PPV-only minimal randomness is not enough (Lo et al., 2024, Schlegel et al., 2022).