When to Switch, Not Just What: Transition Quality Prediction in Clash Royale

Heeyun Heo, Huy Kang Kim · 2026

Read When to Switch, Not Just What: Transition Quality Prediction in Clash Royale by Heeyun Heo, Huy Kang Kim (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-21868.txt.

Introduction

Heeyun Heo and Huy Kang Kim study when to switch decks in Clash Royale, using over 926,000 matches from 34,619 players. Switching frequency runs inversely to win rate across players. This traces to a Zero Switching Cost Assumption in prior recommenders that ignore switching costs and individual propensity. Their Transition Quality Predictor runs Who, When, What. PersonaGate holds back picks for strong players. TimingGate finds when switching beats staying. ScoreFusion ranks candidates by adoptability and transition quality. It reaches a SwitchGap of plus 10.4 percentage points at a 5.4 percent recommendation rate.

Catalog record

CREATORHeeyun Heo, Huy Kang Kim
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42239
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.21868

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

When to Switch, Not Just What: Transition Quality Prediction in Clash Royale Heeyun Heo, Huy Kang Kim Abstract In competitive games, players frequently switch strategies after losing streaks, yet our analysis of 926,334 match records from 34,619 Clash Royale players reveals a counterintuitive pattern: switching frequency is inversely associated with the win rate, with effects that vary substantially across players and situational contexts. We attribute this to a limitation common in many prior recommendation systems, which evaluate strategies by expected quality while overlooking the behavioral cost of switching and individual differences in switching propensity. We refer to this implicit premise as the Zero Switching Cost Assumption . To address this, we reformulate strategy recommendation as a transition-level decision problem and instantiate it as TQP (Transition Quality Predictor), a three-stage pipeline structured as Who → \rightarrow When → \rightarrow What . PersonaGate suppresses recommendations for players whose strategic consistency is empirically associated with superior outcomes. TimingGate identifies moments when switching is likely to yield a net benefit over staying, using a subtype- and state-matched baseline to control for natural win-rate recovery. ScoreFusion ranks candidate strategies by combining an adoptability signal with predicted transition quality ( Δ ​ W ​ R \Delta WR ). We further introduce SwitchGap, an evaluation metric that measures a policy’s discriminative quality without treating observed player choices as optimal ground truth. This property is particularly important because the most frequent switchers record the lowest win rates. The full pipeline achieves a SwitchGap of + 10.4 +10.4 %p at a recommendation rate of 5.4 % 5.4\% , and loss-triggered switchers, despite being the lowest-performing group, benefit the most from subtype-conditioned guidance. K eywords game analytics \cdot player modeling \cdot behavioral heterogeneity \cdot strategy switching \cdot transition quality prediction 1 Introduction In competitive online games, players often switch strategies after a sequence of losses. Prior work has shown that consecutive defeats often trigger coping behaviors such as strategy switching, character switching, and disengagement [ 17 , 3 ] . In meta-volatile environments like Clash Royale , where frequent balance updates reshape the competitive landscape, exploring new strategies can appear to be a rational response. Yet our behavioral log analysis reveals a counterintuitive pattern: the most frequent switchers tend to achieve lower average win rates than those who maintain strategic consistency. Moreover, the magnitude of this effect varies markedly across players’ behavioral profiles, suggesting that who receives a transition recommendation and when it is made are likely key factors. We interpret this heterogeneity through the lens of Adaptation Cost , a construct we use to describe two primary burdens we associate with strategy switching. Motivated by prior evidence that game-specific expertise materially affects match outcomes [ 4 ] , we define Mastery Loss as the transient performance degradation that may arise before a player becomes proficient with a newly adopted strategy. In this study, this possibility is addressed implicitly through the net transition effect formulation, which controls for expected win-rate recovery under matched conditions. We define Tilt-like Reactivity as a set of behaviorally observable reactive responses in post-loss contexts, including abrupt switching and other erratic or non-deliberate adjustments. While this concept is broadly consistent with prior descriptions of tilt, our formulation infers it from behavioral signals rather than direct psychological measurement [ 17 , 9 , 10 , 3 ] . Consistent with this interpretation, the win-rate gaps observed across behavioral groups suggest that the same switching behavior can lead to markedly different outcomes depending on a player’s subtype. Despite the significance of this pattern, prior research leaves two important gaps. First, much prior work has focused on diagnosis—that is, retrospective behavioral classification—rather than prescription: actionable guidance at the moment a decision must be made. Second, many recommendation systems do not explicitly model the intrinsic costs of switching. When these costs are overlooked, systems may recommend transitions that appear theoretically attractive but are poorly timed for the current player. In adjacent recommender settings, unsatisfactory recommendations have been shown to increase churn risk under uncertainty [ 8 ] . We refer to this implicit premise as the Zero Switching Cost Assumption : the assumption that any player can transition instantaneously and without penalty, regardless of mastery degradation, loss-induced reactivity, or individual switching propensity. The fundamental research question is therefore not merely whether to switch, but for whom and at what moment a transition yields a net improvement. We operationalize this question by comparing the outcome of a transition against the win-rate change observed among matched players who maintained their current strategy, thereby controlling for the recovery that might occur even without a strategic change. To address these limitations, we reformulate strategy recommendation as a transition-level decision problem under behavioral constraints, and instantiate this formulation as TQP ( T ransition Q uality P redictor), a three-stage pipeline structured as Who \rightarrow When \rightarrow What . PersonaGate suppresses recommendations for players whose behavioral consistency is empirically associated with superior outcomes. TimingGate identifies moments when switching is likely to yield a net benefit over staying. ScoreFusion then ranks candidate strategies by combining an adoptability signal with predicted transition quality ( \Delta WR ). 2 Related Work 2.1 Player Behavioral Heterogeneity and Strategy Adaptation Competitive game players exhibit systematic behavioral heterogeneity: individuals differ not only in skill level but also in how they respond to in-game events and adapt their strategies over time. Ingram et al. [ 6 ] demonstrated that trajectory-based behavioral signals can identify distinct play styles through an unsupervised LSTM-autoencoder clustering approach, showing that such heterogeneity is both measurable and persistent across gameplay sessions. Building on this line of work, Ingram et al. [ 7 ] paired each identified style with a style-specific expert policy to generate play-style-conditioned advice, improving simulated player performance when the advice is followed. Subsequent work further reported statistically significant performance gaps across behaviorally defined subgroups [ 5 , 15 ] . More broadly, performance-based clustering studies in competitive games have shown that behavioral groups diverge markedly in their performance trajectories over time [ 12 , 13 ] . Taken together, these findings suggest that behavioral subgroup membership is not merely descriptive of play style, but is also associated with meaningful differences in player outcomes. However, existing research has predominantly examined player performance within relatively stable strategic settings, while giving less attention to the strategy transition decision itself as a primary unit of analysis. Even play-style-conditioned advice frameworks typically determine advice by comparing current behavior with an expert model, without estimating the behavioral cost of acting on it or whether the change is likely to outperform staying. As a result, the possibility that the effectiveness of a transition depends on the player’s behavioral profile, that is, the who behind the switch, has remained underexplored in recommendation design. Under the Adaptation Cost lens introduced in Section 1 , the same transition recommendation may therefore yield substantially different outcomes across behavioral subtypes. 2.2 Implicit Costless-Switching Assumptions in Recommendation Systems Strategy recommendation research has evolved from optimizing team synergies to user profiling based on card preferences and historical win rates [ 1 , 11 ] , with recent frameworks attempting to capture transition patterns through meta-learning or frequency-based profiling [ 16 , 18 ] . Beyond game-specific settings, behavior-aware recommender systems

Read the complete text