Live-stream shopping, where a host sells products on camera while viewers chat, like, tip and buy in real time, is widely described as a world of tipping points. Researchers say familiar viewers respond to the feeling of a crowd differently from newcomers. They say a short stay in a room builds trust while a long one starts to feel like wasted time. Econometrics has a standard tool for claims like these. A panel threshold regression is a regression on repeated observations of the same people that lets a slope change abruptly once some "transition" variable crosses a cut-off, with the cut-off estimated from the data. This proposal points out a flaw in how that tool is usually judged. A threshold is declared real when the model with a break beats a model with a single straight line. The straight line is a weak baseline. The search over which variable to split on, and how many breaks to allow, is a tuning budget that the straight line never receives.
We propose to test threshold claims the way the machine-learning literature learned to test new architectures. The threshold model and a properly tuned flexible alternative receive the same specification-search budget. Both are scored on held-out later days, and uncertainty is propagated through the whole search. We apply this protocol to the first micro-level panel threshold study of live-stream commerce, using the public KuaiLive logs. In two sentences: existing tipping-point findings compare a break against a straight line, which is the econometric version of beating an untuned baseline. We ask which live-commerce tipping points survive when the comparison is a tuned smooth curve with the same search budget, and we calibrate that test on data with known ground truth.
Core Claim
A threshold finding is a model-comparison claim, and it inherits every failure mode of unfair model comparison. The weak-baseline literature shows that reported gains often reflect unequal tuning rather than better structure. Neural Collaborative Filtering vs. Matrix Factorization Revisited showed that a regularized dot-product model beats neural collaborative filtering once it is tuned. When Do Neural Nets Outperform Boosted Trees on Tabular Data? found that light tuning of a boosted tree often matters more than switching architectures. Melis et al. found that a well-tuned LSTM beats more elaborate recurrent cells, and Fantastic Pretraining Optimizers and Where to Find Them found that reported optimizer speedups shrink under fair tuning. These papers motivate the proposal. They establish that the gap between a tuned and an untuned baseline can exceed the gap between methods.
Panel threshold studies have not faced this discipline. They apply Hansen-type fixed-effects threshold regressions to aggregate panels: 46 countries in Impact of the digital trade on lowering carbon emissions in 46 countries, 273 cities in Sci-Tech finance, digital economy and high-quality development of regional economy, construction projects in Threshold effect study on the development of prefabricated buildings, and 20 firms with a dynamic SYS-GMM extension in The Dynamic Impact of Diversification Strategy Based on Enterprise Scale Threshold. All of them test the break against linearity. None tests it against a smooth nonlinear curve, and none reports how many transition variables and regime counts were tried. We supply both missing pieces. We also carry the question into a setting where the hypothesized tipping points are behavioural and the panels are large enough to support out-of-time validation.
To separate "a sharp break" from "a smooth bend," we nest both in one model:
$$y_{it} = \mu_i + \beta_1 x_{it} + (\beta_2 - \beta_1)\, x_{it}\, G\!\left(\frac{q_{it}-\gamma}{s}\right) + \varepsilon_{it}.$$
Here $y_{it}$ is user $i$'s engagement in session $t$, $\mu_i$ is a user fixed effect, $x_{it}$ is the driver, and $q_{it}$ is the transition variable. The parameter $\gamma$ is the threshold location and $s>0$ is the transition width. $G$ is the logistic function, and $\beta_1, \beta_2$ are the slopes well below and well above $\gamma$. As $s \to 0$, the model becomes the standard Hansen single-threshold regression. A large $s$ means the "tipping point" is really a gradual bend. The estimated width and its interval therefore give a direct answer to the concern, raised in the threshold literature, that behavioural transitions may be smooth rather than knife-edge.
Research Program
The substantive hypotheses come from the live-commerce literature, and it defines the application. Deng et al. (see also How Social Presence Influences Consumer Well-Being in Live Video Commerce) report that familiarity moderates how social presence, the sense of being with other people during a purchase, affects consumers. How Live Streaming Changes Shopping Decisions in E-commerce describes an efficiency paradox: demonstrations build trust, but long viewing feels like wasted time. Both findings come from surveys or interviews ($N=240$ survey, $N=16$ interviews in the latter). Neither is tested on behavioural logs.
Recommendation work on live commerce does use behavioural logs, but it treats engagement signals as ordinary model inputs and never tests for regimes. HarmonRank uses continuous neural gating. LSEC-GNN uses a static tripartite graph of users, hosts and items. We turn the qualitative claims into two testable threshold hypotheses on KuaiLive, which logs 23,772 users and 452,621 streamers over 21 days, with room start and end times, watch time, clicks, comments, likes and priced gifts.
The first hypothesis concerns familiarity. The entity is a user–streamer pair and the time index is the session. The transition variable is the number of prior sessions the user had with that streamer. This count is fixed before the session starts, which limits the reverse-causality problem that the threshold literature flags for transition variables. The driver is a proxy for social presence: the rate of comments and likes from other sampled users in the same room during the user's stay. The outcome is whether the user comments or gifts, with gift value as a secondary outcome. Gifts are the closest thing to monetary conversion that KuaiLive observes. Product purchases are not in the public data, and we will say so rather than relabel gifts as purchases.
The second hypothesis concerns elapsed watch time. It uses a discrete-time hazard design: for each minute of a stay, we model the probability of a first comment or gift in the next minute as a function of minutes elapsed, with user fixed effects. Viewers who stay longer are a selected group. We therefore compare the same user across sessions and report results with and without conditioning on eventual stay length. We present this analysis as more fragile than the familiarity analysis.
The contribution is the testing protocol, which has three parts. First, the full specification search is written down in advance: candidate transition variables, one versus two thresholds, and trimming fractions. That search defines the budget $K$. Second, every competitor receives the same $K$ tuning trials. We set $K=30$ Optuna trials, following the per-model budget of McElfresh et al., and we also report wall-clock time because equal trials and equal time are not the same thing when models differ in cost. Third, inference resamples whole users and re-runs the entire search inside each bootstrap draw, so the reported uncertainty covers the choice of specification and not only the choice of $\gamma$.
The data split is by time: days 1–14 for fitting, days 15–17 for tuning, and days 18–21 for testing. This lets us check whether a claimed threshold location stays put when estimated on rolling windows. A related open problem, the cost of threshold grid searches on millions of events, is handled by profiling $\gamma$ on a quantile grid of the transition variable and by fitting the smooth-transition model with gradient methods. Its fitted $\gamma$ then serves as a warm start for the sharp model.
Before we interpret anything on real data, the protocol is calibrated semi-synthetically. We keep KuaiLive's real covariates, fixed-effect structure and heavy-tailed activity, in which 70.7% of streamers have fewer than five interactions. We replace the outcome with one generated from a known truth: no nonlinearity, a smooth bend, a sharp single break, or a double break. This measures how often each testing procedure declares a threshold that is not there, and how often it misses one that is.
Baselines
The naive reference point is a linear fixed-effects model with the same user–streamer effects and no nonlinearity. In current threshold papers this is effectively the only comparison. We keep it so we can show how much of any "threshold gain" disappears once stronger baselines enter. We also run the conventional analysis unchanged, a Hansen single- and double-threshold fixed-effects regression with bootstrap tests against linearity, because that is the method whose conclusions are under examination.
The strongest smooth competitor is a penalized-spline additive model with the same fixed effects. It lets the driver's effect vary smoothly with the transition variable through a tensor-product spline, and it gets 30 trials over smoothing penalties and basis sizes. This is the model a sharp threshold must beat. It captures everything a smooth bend can, and it keeps the same within-user identification. A threshold "counts" only if it meets three conditions. It must match or beat the additive model's out-of-time log-loss on days 18–21 with non-overlapping user-bootstrap intervals. Its estimated width $s$ must have an interval excluding widths comparable to the spacing between observed values of $q$. Its location must be stable across rolling windows.
The strongest predictive competitor is CatBoost, tuned with the same 30-trial budget and given target-encoded user and streamer identifiers plus the side features. CatBoost is the strongest competitor because it had the best average rank, 5.06, across 104 primary datasets in When Do Neural Nets Outperform Boosted Trees on Tabular Data?. It can represent step functions natively, so it will find a break if one exists. We found no reported result on Lacuna for CatBoost, or for any model, on KuaiLive engagement prediction. The KuaiLive summary gives no benchmark numbers. The closest live-commerce figures in the file are an AUC of 0.8581 for LSEC-GNN against 0.8103 for NCF and 0.8440 for GCN-RS on LSEC-Small (LSEC-GNN), and a 27.2-point gain in summed AUC for HarmonRank over multi-objective BCE on TenRec (HarmonRank). These figures come from different data and objectives, and they motivate our tuning discipline rather than serving as targets. We also include a tuned regularized matrix-factorization model in the style of Rendle et al., as a check that the engagement signal is not almost entirely explained by user–streamer affinity.
The ablation that isolates the key ingredient is the smooth-transition model with $s$ free, compared with the same model with $s$ forced to zero. Everything else is held fixed: data, fixed effects and budget. Any difference in fit or stability is then due to sharpness alone.
Why not just…?
Why not just fit a tuned boosted tree and read the tipping point off its partial-dependence curve? This is the most serious alternative, and it is included as a baseline rather than dismissed. It has two weaknesses for inference. A partial-dependence curve from a tree mixes differences between users with the response within a user. A newcomer who engages little and a regular who engages a lot can create an apparent jump in familiarity even when no individual changes behaviour at any point. The curve also comes with no interval for the break's location. If CatBoost with identifier encodings predicts clearly better out of time and its curve for held-fixed users shows no jump, we will report that the familiarity "threshold" is a between-user artefact. That would be a substantive negative result.
Why not just rely on the Hansen bootstrap test, which already accounts for searching over the threshold location? It accounts for searching over $\gamma$ under a null of linearity, which is precisely the weak baseline. It does not ask whether a smooth bend explains the data equally well. It does not account for trying several transition variables and regime counts. The semi-synthetic smooth-bend condition measures this directly. If the standard test flags "significant thresholds" there at rates well above its nominal level while our protocol does not, the critique is established. If both behave well, the extra machinery is unnecessary and we will say so.
Why not just fit a spline model and skip thresholds entirely? If splines win everywhere, that is a legitimate outcome. The proposal is built so it can reach it, since the additive model is the main competitor. The reason to keep the threshold family is that operational decisions in live commerce, such as when to trigger a flash sale or when to stop recommending a room to a viewer, need a location with an interval. The width parameter tells us whether such a location is meaningful.
Why not just re-run the published live-commerce recommenders against tuned baselines, as Rendle and colleagues did for collaborative filtering? That audit is worth doing, but it answers a different question: whether architectures beat baselines, not whether behaviour has regimes. It also depends on LSEC and Kuaishou-ELive data whose public availability the file does not establish. We include the matrix-factorization baseline so our results speak to that question on KuaiLive, but a full replication is out of scope.
Proposed Experiments
The semi-synthetic calibration comes first and is decisive for the method. The protocol is useful only if, on KuaiLive-shaped data with 80% of test cases (users) drawn from generators without a sharp break, it keeps false threshold detections near the nominal rate, it recovers true break locations within their intervals, and it is not much less powerful than the standard test against a real sharp break.
On the real logs, the familiarity hypothesis is the main test. Deng et al.'s moderation claim would have a behavioural tipping point only if the threshold model matches the tuned additive model and CatBoost on out-of-time log-loss and AUC, the estimated width is narrow, and the location is stable across rolling windows. The watch-time hazard analysis is the second test and will be reported with its selection caveat.
Throughout, we report what the conventional analysis would have concluded next to what the budget-matched protocol concludes. The size of that gap is the headline result. It shows how much of the "tipping point" evidence in this style of work reflects a weak baseline rather than behaviour. Because the entire case rests on a single Chinese platform over three weeks, and on gifts and comments rather than purchases, conclusions will be stated as limited to that setting.
References
- Neural Collaborative Filtering vs. Matrix Factorization Revisited
- When Do Neural Nets Outperform Boosted Trees on Tabular Data?
- On the State of the Art of Evaluation in Neural Language Models
- Fantastic Pretraining Optimizers and Where to Find Them
- Impact of the digital trade on lowering carbon emissions in 46 countries
- Sci-Tech finance, digital economy and high-quality development of regional economy: empirical evidence from 273 cities in China
- Threshold effect study on the development of prefabricated buildings for energy conservation and emission reduction in the construction industry
- The Dynamic Impact of Diversification Strategy Based on Enterprise Scale Threshold on the Business Performance of Sports Enterprises
- How Social Presence Influences Consumer Well-Being in Live Video Commerce: The Mediating Role of Shopping Enjoyment and the Moderating Role of Familiarity
- How Live Streaming Changes Shopping Decisions in E-commerce: A Study of Live Streaming Commerce
- HarmonRank: Ranking-aligned Multi-objective Ensemble for Live-streaming E-commerce Recommendation
- Leveraging Tripartite Interaction Information from Live Stream E-Commerce for Improving Product Recommendation
- KuaiLive: A Real-time Interactive Dataset for Live Streaming Recommendation