Predictable Errors Research Articles

Assessing the predictability of five intraocular lens calculation methods in eyes with prior myopic keratorefractive lenticule extraction.

To evaluate and compare the predictability of five methods of intraocular lens (IOL) calculation in eyes with prior keratorefractive lenticule extraction (KLEx) for the treatment of myopia. A retrospective case study included 100 eyes of 52 patients who underwent myopia and myopia with astigmatism treatment with small incision lenticule extraction (SMILE). Preoperative and 3-month postoperative measurements of optical biometry and corneal tomography were obtained. The spherical equivalent of the refractive change induced by surgery was converted to the corneal plane (SMILE-dif). A physically well-defined method was developed in which the same IOL model was implanted before and after SMILE. IOL power was calculated using ray-tracing (RT-Sirius), and several IOL power calculation formulas (Kane, EVO 2.0, Barrett Universal II Formula, Hoffer QST) before surgery. After surgery, IOL power was calculated with RT-Sirius, Kane using Mean Pupil Power at 5.5mm by ray tracing, EVO 2.0 Post Myopic LASIK/PRK, Barrett True K and Hoffer QST Post Myopic LASIK/PRK after surgery. The difference between the refractive error induced by the IOL before and after SMILE in the corneal plane (IOL-dif) was compared with SMILE-dif. The predicted error (PE) was calculated as the difference between SMILE-dif and IOL-dif. The PE obtained was 0.26 ± 0.55 diopters (D), 0.10 ± 0.45 D, 0.40 ± 0.37 D, -0.03 ± 0.36 D, 0.02 ± 0.51 D, with RT-Sirius, Kane, EVO 2.0, Barrett True K, and Hoffer QST respectively. PE was not statistically significantly different between Barrett True K and Hoffer QST, with differences being more homogeneous with Barrett, (variance σ2 = 0,13). The absolute EP obtained with Barrett True K achieved 84% of cases within ± 0.5 D, followed by Kane (72%), Hoffer QST (65%), EVO (61%) and RT-Sirius (59%). Barrett True K formula was the most accurate method for IOL calculation in eyes that had undergone SMILE for the correction of myopia. What is known The literature regarding IOL power calculation after SMILE is sparse, and the methods used to estimate corneal power following LASIK/PRK may not be applicable to SMILE procedures. The most common approach to investigating the predictability of IOL calculation formulas involves a theoretical model encompassing the virtual implantation of an IOL. What is new The Hoffer QST formula, Kane formula using Mean Pupil Power at 5.5mm, EVO 2.0, and Sirius' Ray Tracing software had not been previously evaluated using this approach. The Barrett True K formula was the most accurate method for IOL calculation in eyes that had undergone SMILE for myopia correction, outperforming Ray Tracing.

Transparency improves the accuracy of automation use, but automation confidence information does not.

Increased automation transparency can improve the accuracy of automation use but can lead to increased bias towards agreeing with advice. Information about the automation's confidence in its advice may also increase the predictability of automation errors. We examined the effects of providing automation transparency, automation confidence information, and their potential interacting effect on the accuracy of automation use and other outcomes. An uninhabited vehicle (UV) management task was completed where participants selected the optimal UV to complete missions. Low or high automation transparency was provided, and participants agreed/disagreed with automated advice on each mission. We manipulated between participants whether automated advice was accompanied by confidence information. This information indicated on each trial whether automation was "somewhat" or "highly" confident in its advice. Higher transparency improved the accuracy of automation use, led to faster decisions, lower perceived workload, and increased trust and perceived usability. Providing participant automation confidence information, as compared with not, did not have an overall impact on any outcome variable and did not interact with transparency. Despite no benefit, participants who were provided confidence information did use it. For trials where lower compared to higher confidence information was presented, hit rates decreased, correct rejection rates increased, decision times slowed, and perceived workload increased, all suggestive of decreased reliance on automated advice. Such trial-by-trial shifts in automation use bias and other outcomes were not moderated by transparency. These findings can potentially inform the design of automated decision-support systems that are more understandable by humans in order to optimise human-automation interaction.

Generalizable Error Modeling for Human Data Annotation: Evidence From an Industry-Scale Search Data Annotation Program

Machine learning (ML) and artificial intelligence (AI) systems rely heavily on human-annotated data for training and evaluation. A major challenge in this context is the occurrence of annotation errors, as their effects can degrade model performance. This paper presents a predictive error model trained to detect potential errors in search relevance annotation tasks for three industry-scale ML applications (music streaming, video streaming, and mobile apps). Drawing on data from an extensive search relevance annotation program, we demonstrate that errors can be predicted with moderate model performance (AUC=0.65-0.75) and that model performance generalizes well across applications (i.e., a global, task-agnostic model performs on par with task-specific models). In contrast to past research, which has often focused on predicting annotation labels from task-specific features, our model is trained to predict errors directly from a combination of task features and behavioral features derived from the annotation process, in order to achieve a high degree of generalizability. We demonstrate the usefulness of the model in the context of auditing, where prioritizing tasks with high predicted error probabilities considerably increases the amount of corrected annotation errors (e.g., 40% efficiency gains for the music streaming application). These results highlight that behavioral error detection models can yield considerable improvements in the efficiency and quality of data annotation processes. Our findings reveal critical insights into effective error management in the data annotation process, thereby contributing to the broader field of human-in-the-loop ML.

Credit Cycles, Expectations, and Corporate Investment

Abstract We provide a systematic empirical assessment of the Minsky hypothesis that business fluctuations stem from irrational swings in expectations. Using predictable firm-level forecast errors, we build an aggregate index of irrational expectations and use it to provide three sets of results. First, we show that our index predicts aggregate credit cycles. Next, we show that these predictable credit cycles drive cycles in firm-level debt issuance and investment and similar cycles between financially constrained and unconstrained firms, as Minsky predicts. Finally, we show more pronounced cycles in firm-level financing and investment for firms with ex ante more optimistic expectations. (JEL G31, G32, G40, E32, E44)

Accuracy of intraocular lens power calculation formulae for the Yamane technique of secondary fixation: a systematic review and meta-analysis.

This systematic review and meta-analysis aims to assess the refractive outcomes of the Yamane technique for intrascleral fixation of intraocular lenses (SF-IOL) and compare the predictive ability of the various intraocular lens power calculation formulae commonly used in conjunction with the technique. A literature search was conducted in the Medline, Scopus, and Cochrane Library databases for articles published from January 2014 to May 2023. Studies that met the predetermined inclusion criteria were included and subjected to analysis. The primary outcome evaluated was the refractive predictive error, defined as the difference between predicted refraction and post-operative manifest refraction. Eleven studies met the inclusion criteria, with a cumulative sample size of 615 patients (mean age: 66.6 years). Various IOL formulae were used, with SRK/T being the most frequently adopted formula. The overall mean refractive predictive error for all formulae combined was -0.02 D, which was not statistically significant (p = 0.99). Subgroup analysis for individual formulae also showed no significant difference from predicted error for any formula (p > 0.05). The Yamane technique for SF-IOL shows promising refractive outcomes, and the choice of IOL power calculation formula should be tailored based on patient characteristics and surgeon preference. No formula demonstrated superior predictive ability over others. Further research is needed to develop formulae specifically for eyes with secondary aphakia and poor capsular support.

Inflation Expectations with Finite Horizon Planning

Under finite horizon planning, households and firms evaluate a full set of state-contingent paths along which the economy might evolve out to a finite horizon but have limited ability to process events beyond that horizon. We show--analytically and empirically--that such a model accounts for an initial underreaction and subsequent overreaction of inflation forecasts. A planning horizon of four quarters can account for the evidence on the predictability of inflation forecast errors and macroeconomic data. Our identification and estimation strategies combine full-information methods based on aggregate data with regression-based estimates that directly use inflation expectations data.

Accuracy of intraocular lens formulas in combined phacovitrectomy.

To assess the refractive accuracy of eight intraocular lens (IOL) formulas in eyes that underwent combined phacovitrectomy. A retrospective chart review of 59 eyes that underwent uncomplicated phacovitrectomy between 2017 and 2020 at the Johns Hopkins Wilmer Eye Institute. Inclusion criteria were postoperative best corrected visual acuity of 20/40 or better within 6months of surgery and IOL implantation in the capsular bag. The Barrett Universal II (BUII), Emmetropia Verifying Optical (EVOv2.0), Hill-Radial Basis Function (Hill-RBFv3.0), Hoffer Q, Holladay I, Kane, Ladas Super Formula (LSF), and SRK/T formulas were compared for accuracy in predicting postoperative spherical equivalents (SE) using Wilcoxon rank sum tests. Pearson's correlation coefficients were used to assess correlations between biometric parameters and errors for all formulas. Prediction errors of SE ranged from -1.69 to 1.43 diopters (D), mean absolute errors (MAE) ranged from 0.39 to 0.47 D, and median absolute errors (MedAE) ranged from 0.23 to 0.37 D among all formulas. The BUII had the lowest mean error (-0.043), MAE (0.39) and MedAE (0.23). The BUII also had the highest percentage of eyes with predicted error within ± 0.25 D (51%) and ± 0.50 D (83%). Based on MedAE however, no pairwise comparisons resulted in statistically significant differences. Axial length (AL) was positively correlated with the error from the Hoffer Q and Holladay I formulas (correlation coefficients = 0.34, 0.30, p values < 0.01, 0.02 respectively). While all eight IOL formulas had comparable accuracy in predicting refractive outcomes in eyes undergoing combined phacovitrectomy, the BUII and Kane formulas had a tendency to greater accuracy.

Concurrent Encoding of Sequence Predictability and Event-Evoked Prediction Error in Unfolding Auditory Patterns.

Human listeners possess an innate capacity to discern patterns within rapidly unfolding sensory input. Core questions, guiding ongoing research, focus on the mechanisms through which these representations are acquired and whether the brain prioritizes or suppresses predictable sensory signals. Previous work, using fast auditory sequences (tone-pips presented at a rate of 20 Hz), revealed sustained response effects that appear to track the dynamic predictability of the sequence. Here, we extend the investigation to slower sequences (4 Hz), permitting the isolation of responses to individual tones. Stimuli were 50 ms tone-pips, ordered into random (RND) and regular (REG; a repeating pattern of 10 frequencies) sequences; Two timing profiles were created: in "fast" sequences, tone-pips were presented in direct succession (20 Hz); in "slow" sequences, tone-pips were separated by a 200 ms silent gap (4 Hz). Naive participants (N = 22; both sexes) passively listened to these sequences, while brain responses were recorded using magnetoencephalography (MEG). Results unveiled a heightened magnitude of sustained brain responses in REG when compared to RND patterns. This manifested from three tones after the onset of the pattern repetition, even in the context of slower sequences characterized by extended pattern durations (2,500 ms). This observation underscores the remarkable implicit sensitivity of the auditory brain to acoustic regularities. Importantly, brain responses evoked by single tones exhibited the opposite pattern-stronger responses to tones in RND than REG sequences. The demonstration of simultaneous but opposing sustained and evoked response effects reveals concurrent processes that shape the representation of unfolding auditory patterns.

Association of refractive outcome with postoperative anterior chamber depth measured with 3 optical biometers.

This study evaluated the relationship between refractive outcomes and postoperative anterior chamber depth (ACD, measured from corneal epithelium to lens) measured by swept-source optical coherence tomography (SS-OCT), optical low-coherence reflectometry (OLCR), and Scheimpflug devices under the undilated pupil. Patients undergoing cataract phacoemulsification with intraocular lens (IOL) implantation in a hospital setting were enrolled. Postoperative ACD (postACD) was performed with an SS-OCT device, an OLCR device, and a Scheimpflug device at least 1month after cataract surgery. After adjusting the mean predicted error to 0, differences in refractive outcomes were calculated with the Olsen formula using actual postACD measured from 3 devices and predicted value. Overall, this comparative case study included 69 eyes of 69 patients, and postACD measurements were successfully taken using all 3 devices. The postACD measured with the SS-OCT, OLCR, and Scheimpflug devices was 4.59 ± 0.30, 4.50 ± 0.30, and 4.54 ± 0.32mm, respectively. Statistically significant differences in postACD were found among 3 devices (P < 0.001), with intraclass correlation coefficients (ICCs) and Bland-Altman showing good agreement. No significant difference in median absolute error was found with the Olsen formula using actual postACD obtained with 3 devices. Percentage prediction errors were within ± 0.50D in 65% (OLCR), 70% (Scheimpflug), and 67% (SS-OCT) calculated by actual postACD versus 64% by predicted value. Substantial agreement was found in postACD measurements obtained from the SS-OCT, OLCR, and Scheimpflug devices, with a trend toward comparable refractive outcomes in the Olsen formula. Meanwhile, postACD measurements may be potentially superior for the additional enhancement of refractive outcomes.

Improving misspelled word solving for human trafficking detection in online advertising data

&lt;span lang="EN-US"&gt;Social media is used by pimps to advertise their businesses for adult services due to easy accessibility. This requires the potentially computational model for law enforcement authorities to facilitate a detection of human trafficking activities. The machine learning (ML) models used to detect these activities mostly rely on text classification and often omit the correction of misspelled words, resulting in the risk of predictions error. Therefore, an improvement data processing approach is one of strategies to enhance an efficiency of human trafficking detection. This paper presents a novel approach to solving spelling mistakes. The approach is designed to select misspelled words, the replace them with the popular words having the same meaning based on an estimation of the probability of words and context used in human trafficking advertisements. The applicability of the proposed approach was demonstrated with the labeled human trafficking dataset using three classification models: k-nearest neighbor (KNN), naive Bayes (NB), and multilayer perceptron (MLP). The achievement of higher accuracy of the model predictions using the proposed method evidences an improved alert on human trafficking outperforming than the others. The proposed approach shows the potential applicability to other datasets and domains from the online advertisements.&lt;/span&gt;

Sorting a sticky situation: Engaging students in sorting and identifying microfossils from La Brea Tar Pit

Although they do not garner the attention of larger, charismatic fossils, small (1 mm – 1 cm) plant and animal remains are invaluable resources for palaeontologists. These “microfossil” taxa are gener- ally confined to their local habitat and therefore provide insight on the nature of past environments and long-term ecosystem change. For decades, the La Brea Tar Pits (LBTP) in Los Angeles, Califor- nia, USA, has incorporated volunteer sorting of microfossils in their daily research agendas to sup- port scientific discovery in their onsite Fossil Lab. A recent grant-funded research project provided an opportunity to study the Rancho La Brea (RLB) microfossils while extending this “community science” activity beyond the walls of the museum and into local classrooms. From 2018 to 2020, nearly 1500 students contributed by sorting matrix into distinct categories: “plant”, “bone”, “shell”, and “other”. We evaluate the success of this project both in terms of producing usable scientific data and in its impact on the students. We found enjoyment was high among students who completed the ac- tivity. Overall, the level of accuracy of student work was less than that of trained volunteers. However, students were consistent in identifying fossils that were most likely to be valuable for research, and predictable errors provided insight into potential improvements. These data can inform planning for future research needs, lab workflows and opportunities for educational outreach at LBTP and other institutions worldwide.

Remaining useful life prediction of lithium battery based on ACNN-Mogrifier LSTM-MMD

Predicting the remaining useful life (RUL) of lithium batteries is crucial for predicting battery failure and health management. Accurately estimating the RUL allows for timely maintenance and replacement of batteries that pose safety risks. To enhance the safety and reliability of lithium battery operations, this paper proposes a lithium battery life prediction model, attention mechanism-convolutional neural network (ACNN)-Mogrifier long and short-term memory network (LSTM)-maximum mean discrepancy (MMD), based on ACNN, Mogrifier LSTM, and MMD Feature Transfer Learning. Firstly, the capacity degradation data from historical life experiments of lithium batteries in both source and target domains are extracted. The whale optimization algorithm (WOA) is employed to optimize the parameters of variational modal decomposition, enabling the decomposition of the historical capacity degradation data into multiple intrinsic mode functions (IMFs) components. Secondly, highly correlated IMF components are identified using the Pearson correlation coefficient (Pearson) to reconstruct the capacity sequence, which characterizes the capacity degradation information of the lithium batteries. These reconstructed sequences are inputs to the ACNN model to extract features from the capacity degradation data. The extracted features are then utilized to compute MMD values, quantifying the distribution differences between the two domains. The Mogrifier LSTM neural network estimates the capacity values of the source and target domains and calculates the loss functions by comparing them to the actual capacity values. These loss functions, along with the computed MMD values, are combined to obtain the combined loss function of the model. Finally, the ACNN-Mogrifier LSTM-MMD is applied to the target domain data to formulate the lithium battery RUL prediction model. The effectiveness of the proposed method is validated using CACLE and NASA lithium battery datasets, The experimental results demonstrate that the predicted error of the RUL for the B5 battery is less than 6% for mean absolute percentage error (MAPE) and less than 1 for . Similarly, the RUL prediction error for the B6 battery is below 10% for MAPE and less than 1 for . This indicates higher accuracy compared to other prediction methods, along with improved robustness and practicality.

Exchange rates and binary political events

Abstract This article introduces a rational expectations model that explains exchange rate dynamics and the predictability of forecast errors using private (aggregated via order flow) and public (probabilities of a binary event) information. We test the model for the periods leading up to the presidential impeachment vote in Brazil, the Brexit Referendum, and Donald Trump’s election in 2016. Proxies of the physical probabilities of these events reveal that they are a crucial source of pricing information for the BRL, GBP, and MXN currency pairs with the US dollar. They also explain forecast errors. The information content of order flow changes before and after an actual regime change resolves uncertainty.

A probabilistic perspective on predictability of solar irradiance using bootstrapped correlograms and ensemble predictability error growth

Recently in “Correlogram, predictability error growth, and bounds of mean square er-ror of solar irradiance forecasts” [Renew. Sustain. Energy Rev. 167 (2022) 112736], a new measure of predictability, applicable to solar irradiance in particular, was proposed. That predictability measure is expressed as one minus the ratio of the smallest attainable and highest tolerable root mean square errors (RMSEs) for a specific forecasting situation. Whereas the smallest attainable RMSE can be derived from the predictability error growth (PEG) of a numerical weather prediction model, the highest tolerable RMSE is computed from the forecasts of a naïve standard of reference. Despite that proposal has pioneered our understanding on the predictability of solar irradiance, answers to some technical challenges remain opaque. One of those technical challenges is the lack of uncertainty quantification of the predictability estimate, which is what this work seeks to resolve. Two conceptions, namely, ensemble PEG and bootstrapped correlograms, are herein introduced to facilitate the probabilistic representations of the smallest attainable and highest tolerable RMSEs, respectively. It is found that the uncertainty in predictability estimates is very small as compared to the range over which predictability varies across climatic conditions and forecast horizons.

Robust neural decoding for dexterous control of robotic hand kinematics

BackgroundManual dexterity is a fundamental motor skill that allows us to perform complex daily tasks. Neuromuscular injuries, however, can lead to the loss of hand dexterity. Although numerous advanced assistive robotic hands have been developed, we still lack dexterous and continuous control of multiple degrees of freedom in real-time. In this study, we developed an efficient and robust neural decoding approach that can continuously decode intended finger dynamic movements for real-time control of a prosthetic hand. MethodsHigh-density electromyogram (HD-EMG) signals were obtained from the extrinsic finger flexor and extensor muscles, while participants performed either single-finger or multi-finger flexion-extension movements. We implemented a deep learning-based neural network approach to learn the mapping from HD-EMG features to finger-specific population motoneuron firing frequency (i.e., neural-drive signals). The neural-drive signals reflected motor commands specific to individual fingers. The predicted neural-drive signals were then used to continuously control the fingers (index, middle, and ring) of a prosthetic hand in real-time. ResultsOur developed neural-drive decoder could consistently and accurately predict joint angles with significantly lower prediction errors across single-finger and multi-finger tasks, compared with a deep learning model directly trained on finger force signals and the conventional EMG-amplitude estimate. The decoder performance was stable over time and was robust to variations of the EMG signals. The decoder also demonstrated a substantially better finger separation with minimal predicted error of joint angle in the unintended fingers. ConclusionsThis neural decoding technique offers a novel and efficient neural-machine interface that can consistently predict robotic finger kinematics with high accuracy, which can enable dexterous control of assistive robotic hands.

A Comprehensive Empirical Evaluation of Biases in Expectation Formation

We revisit predictability of forecast errors in macroeconomic survey data, which is often taken as evidence of behavioral biases at odds with rational expectations. We argue that to reject rational expectations, one must be able to predict forecast errors out of sample. However, the regressions used in the literature often perform poorly out of sample. The models seem unstable and could not have helped to improve forecasts with access only to available information. We do find some notable exceptions to this finding, in particular mean bias in interest rate forecasts, that survive our out-of-sample tests. Our findings help narrow down the set of biases that merit closer attention of researchers in behavioral macroeconomics.

Technology Focus: Acidizing (June 2023)

The first acid job dates to 1895. Challenges remain, however, when it comes to lithology, coverage, penetration, and management of reaction products. These themes remain a focus for stimulation engineers. In paper SPE 204879, for example, a novel in-situ foam-generation system is introduced that provides effective diversion and leaves no residue. Field applications suggest significantly better performance compared with other diversion systems. In paper SPE 203985, the authors present a 3D acid-fracturing model to calculate fracture geometry induced through multistage alternating injection of pad and acid fluids. Alternating acid with high-viscosity polymer pad promotes nonuniform acid etching and viscous fingering, which leads to longer and narrower channels. Concurrently, investigators are studying methods to achieve effective near-wellbore and far-field bridging using a combination of solids and fiber-laden acid. For those wishing to delve deeper into this subject, paper SPE 206255 makes interesting reading. Completions can play a role in improving acid treatments. Limited entry, as presented in paper SPE 211394, and the so-called “segmented and selective” completions are at the forefront. Paper IPTC 23003 introduces the concept of sweet spots, which are then completed selectively with packers and sliding sleeves. Paper SPE 213406 describes a relatively complex architecture for a trilateral TAML Level 2 well with openhole packers and ball-activated sleeves. Thus, it can be said that most enabling technology is out there. Lastly, engineers have begun to leverage benefits that big data and artificial intelligence potentially offer. Predicting an outcome of an acid job remains hit-and-miss at best—a gap engineers must bridge. In paper SPE 208172, the authors present a predictive model to estimate the efficiency of acid-fracturing treatments in naturally fractured reservoirs using an artificial neural network. They report a predictability error of less than 10%. This may be what we are looking for. I hope you find the selections enlightening. Recommended additional reading at OnePetro: www.onepetro.org. SPE 207514 Subsea Multiwell High-Flow-Rate Riserless Acid Stimulation Campaign With Two-Vessel Approach: A Comparative Case Study by Irma Kusumawati, ConocoPhillips, et al. SPE 205599 A New Method of Carbonate Matrix Stimulation Modeling and Optimization Using Field Experimental Data for Acid Efficiency Calibration and Case Studies in Kazakhstan by Ruslan Kalabayev, SLB, et al. SPE 209673 Reducing Carbon Footprint of Matrix Acidizing in Carbonate Formations: How Much Acid Do We Really Need To Pump? by Mahmoud T. Ali, Baker Hughes

Reversible Data Hiding in Encrypted Image Based on Bit-Plane Redundancy of Prediction Error

In this paper, we propose a reversible data hiding scheme in an encrypted image based on bit-plane redundancy of prediction error. The scheme greatly improves the embedding capacity while maintaining lossless image recovery and error-free secret data extraction. Firstly, the original image is preprocessed to obtain the prediction error image. After the error matrix is divided into blocks, the corresponding block type is obtained. Secondly, the predicted error image is encrypted with stream cipher and the encryption matrix blocks are scrambled to ensure the security of the scheme. Finally, after embedding the block type value into the encrypted image, the spare room corresponding to each block was obtained, which was used to embed the secret data. The scheme makes full use of the spatial correlation of the pixels in the block, so it improves the embedding rate. By selecting 100 images in each dataset of BOSSbase and BOWS-2, when the block size is 3×3, the average embedding rate of our scheme can reach 3.56 bpp and 3.81 bpp, respectively. The performance of the proposed method is better than the other schemes with similar properties.

Imperfect Exchange Rate Expectations

Abstract Using survey data, we document that predictable exchange rate forecast errors are responsible for the uncovered-interest-parity (UIP) puzzle and its reversal at longer horizons. We develop a general-equilibrium model based on shock misperception and over-extrapolative beliefs that reconciles these and other major exchange rate puzzles. These beliefs distortions generate both under- and over-reaction of expectations that account for the predictability of forecast errors about interest rates, exchange rates, and other macroeconomic indicators. In the model, forecast errors are endogenous to monetary policy and explain the change in the behavior of UIP deviations that emerged after the global financial crisis.

The Role of Ocean–Atmosphere Interactions in Tropical Cyclone Intensity Predictability

Abstract The negative feedback between tropical cyclone (TC) intensity and sea surface temperature (SST) plays an important role in TC development. In this study, ocean–atmosphere coupled and uncoupled ensemble forecasts are conducted to investigate the dynamics of error growth and predictability of TC intensity in an ocean–atmosphere coupled system. For the TC–ocean coupled system, the TC intensity–SST negative feedback is the essential mechanism to reduce the error growth of TC intensity by two routes, and thereby improves the TC intensity predictability. For the first route (atmosphere-limited route), the TC-induced SST cooling slows the intensification rate of the TC and weakens the final TC intensity, thereby reducing the error growth of TC intensity. In this route, the TC intensity spread is limited by the magnitude of TC intensity, while SST can be regarded as an environmental forcing. For the second route (atmosphere–ocean mutually influenced route), the interaction between the TC intensity spread and SST spread is dominant. The increasing TC intensity spread could lead to an increase in SST cooling spread, and then reduce the TC intensity spread through the negative feedback. In other words, the more (less) intense TC produces stronger (weaker) SST cooling, and thereby limits (enhances) further TC intensification in an ensemble forecast. In the second route, initial ocean temperature uncertainty could suppress the TC intensity spread reduction. Significance Statement Tropical cyclones force the sea surface and can lead to its cooling. This cooled sea surface can then suppress tropical cyclone intensification. The purpose of this study is to better examine the influence of such an interaction between a tropical cyclone and the ocean on tropical cyclone forecasts. We explore how accurately representing the interaction can improve the capacity to forecast tropical cyclone intensity. Given that many weather forecasting centers have considered this interaction in their models, this study should help them to understand and improve their forecasts.