{"id":"30ba1597-c89f-48db-983a-141cf197c727","arxiv_id":"1908.08127","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"The authors forecast 75K daily e-scooter trips in Manhattan from a three-city regression and estimate that carpool, bike, and taxi trips are the most likely modes to be replaced.","lead":"This paper predicts that 2,000 e-scooters in Manhattan would generate about 75,000 daily trips, then uses a statistical model to break those trips down by which existing modes they might replace. The authors estimate the largest substitutions would come from carpool, bike, and taxi trips, with annual revenue near $77 million.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 75K daily trip forecast implies 37.5 trips per scooter per day in Manhattan, roughly 13x the utilization of the pilot cities used to estimate Eq. 1; this suggests the central demand forecast is not credible.","rationale":"The reader's weakest assumption focused on transferability of the log-log trip generation model from three pilot cities to Manhattan. That is a valid concern, but the more pointed problem is that the resulting forecast implies a per-scooter utilization rate that is an order of magnitude above the utilization in the same pilot cities the model was estimated on. Even if every elasticity were transferable, a 2,000-scooter fleet generating 75K daily trips is not a credible scenario under normal e-scooter operations: at 12 minutes per trip, each scooter would need to be ridden for 7.5 hours per day, with no time lost to charging, maintenance, or repositioning. This makes the central number in the abstract and the revenue estimate derived from it unreliable. The substitution shares are fitted to this same forecast, so they are conditional on a demand number that fails a basic sanity check. The manuscript is transparent about its methods and explicitly labels some components exploratory, but transparency does not repair an implausible headline forecast. A condition could not fix this without changing the fleet-size scenario or the forecast itself, so the appropriate adjustment is to reject the central claim as stated. The route-based and substitution-modeling ideas may still be worth pursuing, but the paper's main quantitative conclusions should not be accepted in their current form.","tokens_in":16572,"tokens_out":8884,"duration_ms":93161,"concrete_test":"Compute daily trips per scooter in the estimation cities using the totals reported in Section 3.1: Portland, 700K trips over roughly 122 days with 2K scooters, about 2.9 trips/scooter/day; Chicago, 821K trips with 2.5K scooters, about 2.1-2.7 trips/scooter/day. Compare with the Manhattan forecast of 75K daily trips divided by 2K scooters, i.e., 37.5 trips/scooter/day. If the forecast exceeds the observed pilot-city per-scooter rate by more than a factor of 5, then the 2,000-scooter scenario fails a basic fleet-productivity bound. A second check would be to back-solve Eq. 1 for the fleet size needed to produce 75K daily trips under Manhattan demographics; if the required fleet exceeds 10K, the headline 2K-scenario forecast is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central quantitative claim is the 75K daily e-scooter trip forecast for a 2,000-scooter Manhattan deployment (Section 5.1), which drives the $77M revenue estimate (Section 5.4) and the substitution shares in the abstract. That forecast can be checked directly against the paper's own pilot data. Portland reported 700K trips over a four-month pilot with 2K scooters (Sections 3.1 and 4.1), which is about 5.7K trips/day, or 2.9 trips per scooter per day. Chicago reported 821K trips with 2.5K scooters, roughly 2.1-2.7 trips per scooter per day. The Manhattan forecast of 75K daily trips with 2K scooters is 37.5 trips per scooter per day, 13-18x the utilization observed in the very cities used to fit Eq. 1. With the paper's own assumption of 12-minute average trips, this requires 7.5 hours of riding per scooter per day before rebalancing, charging, and deadheading, which is not plausible for a shared dockless fleet. This is not just a transferability concern; it is an internal consistency check on the headline number. Eq. 1 is a log-log model with a negative Land Area elasticity and a city-level scooter count, so applying it to Manhattan's small, dense zip codes extrapolates far outside the estimation range. The per-scooter productivity is the clearest symptom. Because the multifactor substitution model is fitted in-sample to the same 75K forecast, its shares inherit the implausibility.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a two-stage model to forecast e-scooter demand in Manhattan and attribute that demand to substituted travel modes. In the first stage, a log-log regression (Eq. 1) is estimated on 48 zip-code observations from Portland, Austin, and Chicago, using population×age ratio, land area, and fleet size as covariates. Applying this model to Manhattan with an assumed 2,000-scooter fleet yields a forecast of 75,000 daily e-scooter trips (Section 5.1). The second stage fits a nonlinear multifactor model (Eq. 4) that decomposes these 75,000 trips into direct substitutions from bike, walk, carpool, taxi, auto, and public transit, plus access/egress substitutions to transit, as a function of distance. The model reports, for example, that e-scooters could replace 32% of carpool trips, 13% of bike trips, and 7.2% of taxi trips, and it projects $77 million in annual revenue (Section 5.4). The paper positions these as the first multi-city e-scooter demand forecast with a fleet-size elasticity and a distance-based mode-substitution decomposition.","tokens_in":16982,"tokens_out":6027,"duration_ms":56785,"significance":"If the forecast were reliable, the paper would provide a useful planning tool for cities considering e-scooter programs and would extend the micromobility literature by explicitly modeling substitution from multiple modes and access/egress trips. The authors are to be credited for assembling multi-city pilot data, including a fleet-size variable, and for providing a bootstrap procedure for the second-stage parameters. However, the central quantitative claims are not currently supported: the first-stage model explains only 31% of variance, the forecast implies per-scooter utilization an order of magnitude above the pilot cities used for estimation, and the second-stage model is fitted to the same forecast it is supposed to explain. As a result, the headline numbers (75K daily trips, $77M revenue, and the substitution shares) are not credible in their present form.","major_comments":[{"comment":"The first-stage trip generation model is too weak to support the 75K daily forecast. The regression has R²=0.314 (adjusted 0.267) from only 48 training records, and the 'Scooters' variable is constant within each city, so its coefficient (0.7812) cannot be separated from city-specific unobservables such as pilot program design, regulatory environment, or marketing effort. The negative land-area elasticity (−1.108) then extrapolates to Manhattan's much smaller zip codes (average 0.5 sq mi vs 9.71 in Portland, 18.51 in Austin, 4.12 in Chicago; Table 1), far outside the estimation range. The out-of-sample validation in Section 4.2 is limited to 12 records and a single Hoboken observation with an error of 0.42, which is not sufficient to establish transferability. The 75K point forecast therefore rests on an extrapolation from a poorly identified, weakly fitting model.","section":"§4.1, Eq. (1), Table 5"},{"comment":"The headline forecast of 75K daily e-scooter trips with 2,000 scooters implies 37.5 trips per scooter per day. The paper's own pilot data give 2.9 trips per scooter per day in Portland (700K trips over about four months with 2K scooters, Section 3.1) and roughly 2.1–2.7 in Chicago (821K trips with 2.5K scooters), so the Manhattan forecast is 13–18 times the utilization observed in the estimation cities. With the paper's assumption of 12-minute average trips (Section 5.4), this requires 7.5 hours of riding per scooter per day before rebalancing, charging, and deadheading, which is not plausible for a shared dockless fleet. This is an internal inconsistency, not merely a transferability caveat: the model applied to Manhattan produces a per-scooter productivity that contradicts the very data used to estimate it. The $77M annual revenue estimate inherits this implausibility.","section":"§5.1 and §5.4"},{"comment":"The multifactor model uses the forecasted 75K trips as its dependent variable, so the reported substitution shares (32% carpool, 13% bike, 7.2% taxi) are in-sample fitted quantities that mechanically reproduce the first-stage forecast; they are not independent predictions. The bootstrap procedure resamples the 318 TAZ observations conditional on the fixed forecast, so it does not propagate the substantial first-stage estimation error. Moreover, the constant C=73.365 alone contributes about 23K trips (Section 5.2), meaning a large share of the decomposition is attributed to an unmodeled constant rather than to mode substitution. Several mode factors are not statistically significant (F_walk t=0.350, F_citi_bike t=1.093, F_auto t=1.677), which weakens the conclusion that carpool, bike, and taxi are the dominant substituted modes. These issues undermine the second-stage claims as presented.","section":"§4.4, Eq. (4), Table 6"},{"comment":"The distance-decay relationship P_d = β_δ / δ_d is imposed by assumption rather than estimated from the data; only the single parameter β_δ is calibrated. Consequently, the claim that the substitution probability 'drops by an order of magnitude' from 0.5 to 5.5 miles (Table 7) follows mechanically from the chosen functional form and is not a data-driven finding. The paper should at least test alternative functional forms or acknowledge that this structure is a modeling assumption, and it should be cautious in presenting the distance-based substitution shares as empirical results.","section":"§4.4, Eq. (2), Table 7"}],"minor_comments":[{"comment":"The text refers to 'Mean Absolute Error' while the figure caption says 'Mean Absolute Deviation'; please make the terminology consistent.","section":"§4.2, Figure 5"},{"comment":"The city of Austin, Texas is referred to as 'Texas' in several places (e.g., 'Texas data points have relatively higher error bound' and the Table 1 header). Please use 'Austin' consistently.","section":"§4.2 and Table 1"},{"comment":"Figure 4(a) is described as a q-q plot of e-scooter ridership in Portland, but the model is estimated on pooled data from three cities; please clarify which residuals or data are shown in the figure.","section":"Figure 4"},{"comment":"The notation for the access-time coefficient is inconsistent: Eq. (6) uses β_5 while Table 6 reports β_6=0.493 and also lists β_7, β_8, β_9 with N/A t-stats. Please align the subscripts and explain the zero-valued coefficients.","section":"Eq. (6) and Table 6"},{"comment":"The paper states in Section 4 that the model 'is not meant to make behavioral predictions of which modes would be substituted by e-scooter,' yet the abstract and Section 5.2 present 'could replace 32% of carpool, 13% of bike, and 7.2% of taxi trips.' This tension between the stated scope and the reported conclusions should be resolved explicitly.","section":"§4, §5.2, and Abstract"},{"comment":"The text says 'All variables are substantially significant,' but the constant has p=0.0949 (significant only at the 10% level); please phrase this more precisely.","section":"§4.1, Table 5"}],"recommendation":"reject","confidential_remarks":"The paper has a useful research question and assembles a dataset that could support a pilot-level analysis, but the central forecast fails an internal consistency check (per-scooter utilization is 13–18 times the pilot cities' values) and the second-stage model is circular with respect to the first-stage forecast. These are load-bearing issues that cannot be fixed by local edits. I would encourage the authors to revisit the first-stage model specification, validate against actual trip data from Manhattan or a comparable dense urban area, and either propagate uncertainty or substantially temper the quantitative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is the first study I've seen that pools e-scooter pilot data from several cities to estimate a trip-generation model with fleet size as a variable, and then decomposes the forecasted demand into direct versus access/egress trips by mode and distance. That framing is genuinely useful for micromobility planning. Second, the headline forecast – 75K daily trips from 2,000 scooters in Manhattan – is not credible as a point estimate. It implies 37.5 trips per scooter per day, about 13 times the utilization observed in Portland (2.9) and Chicago (2.7), the very cities used to estimate the model. That is not a minor concern; it's an internal consistency check that fails hard.\n\nWhat the paper does well: it brings in fleet size variation across cities, which is a real improvement over single-city models. The distance-based impedance parameter and the split between competitive and complementary substitution is new for e-scooters and gives the work a policy-relevant output: where and which modes scooters might draw from. The authors are also honest in the text – they explicitly call the multifactor model exploratory and say it is not meant for behavioral prediction. That disclaimers are better than the abstract, which sells the substitution shares as estimates rather than scenario illustrations.\n\nWhere the paper is soft, in proportion: the first-stage regression has R^2 of 0.31 on 48 records, and the fleet-size variable is a city-level constant, so it's confounded with any city-specific unobservable. There is no uncertainty propagation through the two-stage chain, and the substitution model is fitted to the very forecast it explains, so the 32% carpool and 7.2% taxi figures are in-sample fitted quantities, not independent predictions. The data cleaning step that drops zip codes with fewer than 10 daily trips adds selection bias. The Hoboken validation, which they call satisfactory, is a 42% error on a single point. None of this is fatal to the paper's value as a transparent, reproducible-looking exercise, but it does mean the numbers should be read as scenario output, not forecasts.\n\nWho is this for? Transit planners, micromobility researchers, and anyone doing NYC market sizing. It deserves a serious referee, not a desk reject, because the framework is new and the flaws are identifiable and fixable. In review, I'd ask the authors to address the per-scooter utilization check, propagate uncertainty, and reframe the substitution shares as illustrative. I'd want to see the script or data to verify the aggregation from zip to TAZ. If they can do that, the paper could be a legitimate reference for e-scooter demand modeling.\n\nMy recommendation: send it to peer review with major revisions. It's not ready as-is, but the core idea is worth the referee time.","headline":"The paper's 75K daily Manhattan e-scooter forecast is internally inconsistent with the pilot data used to fit the model, but the multi-city demand framework and the distance-based mode-substitution decomposition are genuinely new and worth engaging with.","tokens_in":17465,"tokens_out":2409,"would_cite":false,"duration_ms":25599,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"E-scooter substitution of Manhattan trips is forecast at 75,000 daily rides, mostly replacing short carpool, bike, and taxi trips.","keywords":["e-scooter demand forecasting","micromobility","mode substitution","trip generation model","nonlinear multifactor model","Manhattan","dockless scooters","access/egress trips"],"falsifier":"A direct test is to compare the forecast with realized operations: a Manhattan e-scooter program capped at 2,000 vehicles should show roughly 75,000 daily trips; if the observed count falls outside the model's out-of-sample error range (about ±20% on log trips, coefficient of variation 0.27), the transferability assumption fails. A second test is to survey riders' prior mode and check whether the shares replacing carpool, bike, and taxi are close to 32%, 13%, and 7.2%.","tokens_in":16338,"feed_emoji":"🛴","tokens_out":11199,"duration_ms":93217,"temperature":0.7,"pith_summary":"This paper forecasts how many e-scooter trips Manhattan would generate if 2,000 dockless scooters were deployed, and which existing trips those rides would replace. It estimates a log-log trip-generation model from zip-code-level e-scooter ridership in Portland, Austin, and Chicago, then applies that model to Manhattan's zones to get 75,000 daily trips, about 1% of all intra-Manhattan travel and roughly 60% more than Citi Bike's daily ridership. A nonlinear multifactor model then attributes those trips statistically to existing modes: up to 32% of carpool trips, 13% of bike trips, 7.2% of taxi trips, and small shares of walking and auto trips, with substitution concentrated at short distances. The paper also finds that e-scooters would substitute for some access/egress trips to public transit, and estimates $77 million in annual revenue under a $1 plus $0.15 per minute fare. If correct, the forecast gives cities and operators a first quantitative picture of the latent micromobility market and the modes it draws from.","feed_headline":"E-scooters would draw 75,000 daily Manhattan trips","feed_subtitle":"2,000 scooters would replace 32% of carpool, 13% of bike, and 7.2% of taxi trips.","key_machinery":"The machinery is a two-stage statistical decomposition. Stage one is a log-log trip-generation regression (Eq. 1) that converts zip-code demographics and a citywide fleet-size variable into predicted e-scooter trips per zone. Stage two is the nonlinear multifactor model (Eq. 4): it treats the predicted trips as the dependent variable and, as independent variables, observed trips by mode and distance from Manhattan's travel survey, multiplied by two parameter sets — $F_m$, the fraction of each mode's trips that e-scooters could replace (fixed across distance), and $P_d = \\beta_d / \\delta_d$, a distance-decay competition probability that shrinks as the average trip distance $\\delta_d$ grows. A separate access-trip factor $F_{transit,i} = \\beta_5 t_i^{access} + \\beta_6 t_i^{egress}$ captures e-scooter substitution for first- and last-mile transit access. Fitting both parameter sets by least squares, with bootstrap confidence intervals, reveals which modes are statistically similar to e-scooter trips and how that similarity decays with distance.","core_discovery":"The paper's central claim is that Manhattan has a large latent e-scooter market that can be quantified even before local ridership data exist. Using a log-log regression\n$$\\ln R_i = \\beta_0 + \\beta_P \\ln(\\text{Population}_i \\times \\text{AgeRatio}_i) + \\beta_L \\ln(\\text{LandArea}_i) + \\beta_S \\ln(\\text{Scooters}) + \\varepsilon_i$$\nestimated on 60 zip-code observations from three pilot cities, the paper predicts 75,000 daily e-scooter trips for a 2,000-scooter Manhattan fleet. A second, nonlinear multifactor model\n$$R_{esco,i} = C + \\sum_m F_m \\sum_d P_d N_{m,i,d} + \\sum_d (1-P_d) F_{transit,i} N_{transit,i,d} + \\gamma_i$$\nwith $P_d = \\beta_d / \\delta_d$ attributes those trips to existing modes, finding statistically significant substitution from carpool ($F=0.636$), bike ($F=0.226$), and taxi ($F=0.184$), with the distance-competition term falling from 0.986 at 0.5 miles to 0.09 at 5.5 miles. The model also estimates an access/egress factor for public transit, $F_{transit,i} = \\beta_5 t_i^{access}$, implying e-scooters can replace a share of first- and last-mile transit access trips. Taken together, the paper claims e-scooters would systematically replace short carpool, bike, taxi, and transit-access trips rather than merely add new travel.","pith_inferences":["A testable extension is to apply the same two-stage structure to another city with pilot data; the fleet-size coefficient of 0.781 implies regulators can use scooter caps to tune total trip volume, not just placement.","The strong distance decay suggests operators could price by distance rather than by time, or concentrate parking and rebalancing on sub-2-mile trips, to flatten the revenue curve.","Since transit-access substitution appears nearly distance-flat, transit agencies could treat e-scooters as a stable first- and last-mile feeder whose revenue does not erode on longer transit trips.","Because the fleet-size effect is identified from only three city-level observations, the exact 75,000 figure is conditional on that comparison; the modal substitution shares are the more portable insight."],"forward_implications":["A 2,000-scooter Manhattan fleet would generate about 75,000 daily trips and about $77 million in annual revenue at a $1 plus $0.15 per minute fare.","E-scooter competition is a short-distance phenomenon: the competition probability drops from 0.986 at 0.5 miles to 0.09 at 5.5 miles, so the mode would serve a last-mile niche rather than long trips.","Carpool is the most exposed mode, with up to 32% of carpool trips replaceable, followed by bike at 13% and taxi at 7.2%; walking and auto substitution are small.","E-scooters would draw not only direct trips but also access/egress trips to public transit, and revenue from that access segment is nearly flat with distance, unlike direct-trip revenue.","Because e-scooters are dockless, they would reach Manhattan areas that station-based Citi Bike does not cover, helping explain the forecast's 60% premium over Citi Bike ridership."],"supporting_citations":[{"why":"Supplies Portland's pilot ridership and rider-survey data used to estimate the trip-generation model and substitution shares.","marker":"(PBOT, 2018)"},{"why":"Supplies Austin's pilot ridership and survey data, one of the three estimation cities.","marker":"(CATD, 2019)"},{"why":"Supplies Chicago's pilot ridership data used in the trip-generation regression.","marker":"(CDOT, 2020)"},{"why":"Provides the zip-code demographics that enter as independent variables in the demand model.","marker":"(ACS, 2018)"},{"why":"Provides the Manhattan trips by mode and distance that form the independent variables of the multifactor substitution model.","marker":"(NYMTC, 2011)"},{"why":"Supplies the synthetic population and validates the mode shares used to represent Manhattan's trip landscape.","marker":"He et al., 2020"},{"why":"Provides daily Citi Bike ridership used as a comparison mode and as a direct substitute in the multifactor model.","marker":"(Citi Bike, 2018)"},{"why":"Establishes that e-scooters substitute short private-auto trips, supporting the distance-decay specification.","marker":"(Smith and Schwieterman, 2018)"}],"fun_headline_variants":["E-scooters could replace 32% of carpool, 13% of bike trips in NYC","75K daily e-scooter trips forecast for Manhattan with 2K scooters","Study predicts e-scooters replace carpool, bike, and taxi trips","Manhattan e-scooter model: 75K daily trips, 32% carpool replaced"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that Manhattan riders will respond to scooter supply the way riders in the three pilot cities did; the paper's closing remarks list omitted factors such as weather, trip chaining, and spatial correlation, so a different response in New York would change the 75,000-trip forecast and every substitution share.","fun_headline_variants_meta":{"raw":{"variants":["E-scooters could replace 32% of carpool, 13% of bike trips in NYC","75K daily e-scooter trips forecast for Manhattan with 2K scooters","Study predicts e-scooters replace carpool, bike, and taxi trips","Manhattan e-scooter model: 75K daily trips, 32% carpool replaced"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1825,"prompt_tokens":1097,"completion_tokens":728,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":632}},"tokens_in":713,"tokens_out":728,"duration_ms":7042,"temperature":1.0,"reasoning_tokens":632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:48:47.028559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to compare the forecast with realized operations: a Manhattan e-scooter program capped at 2,000 vehicles should show roughly 75,000 daily trips; if the observed count falls outside the model's out-of-sample error range (about ±20% on log trips, coefficient of variation 0.27), the transferability assumption fails. A second test is to survey riders' prior mode and check whether the shares replacing carpool, bike, and taxi are close to 32%, 13%, and 7.2%.","supporting_citations":[],"review_version":1}