Methodology

Problem definition

A query contains a current time t and taxi zone o. A strategy returns three distinct destination zones in ranked order.

The repository uses 263 TLC zones and 48 half-hour slots per weekday, producing 336 recurring weekly time states.

Temporal protocol

  • Training: 2023-01-01 through 2023-01-24.

  • Public validation: 2023-01-25 through 2023-01-31.

  • A trip is retained only when pickup and dropoff both satisfy the corresponding boundary.

  • Demand, fare, travel time, OD transitions, and duration statistics are built from training data only.

The public validation labels may be used for a pre-declared parameter grid, but final scientific evaluation requires an untouched later test period. The audit utilities provide expanding rolling-window splits for multi-month data.

Data cleaning

The executable rules remove:

  • out-of-boundary or non-positive-duration trips;

  • missing required fields;

  • invalid pickup/dropoff zone IDs;

  • fare outside \(0–\)200;

  • duration outside 1–240 minutes;

  • distance outside 0.1–100 miles when available;

  • implied speed above 80 mph;

  • duplicate trip keys.

The cleaning command can now split the official monthly parquet into chronological train and validation inputs automatically.

Baseline 1: hot zones

For the strictly next half-hour boundary:

\[ U_{B1}(z,t)=D(z,\mathrm{next}(t)). \]

The three largest training pickup counts are returned.

Baseline 2: single-step utility

\[ U_{B2}(o,z,t)=\frac{D(z,t)\bar f(z,t)}{T(o,z)+1}. \]

T is the directed Dijkstra travel-time matrix estimated from training OD durations.

Finite-horizon continuation model

Pickup probability is a scenario model, not an empirically identified probability:

\[ p(s,z)=\frac{D(s,z)}{D(s,z)+240}. \]

The terminal one-step value is

\[ V_1(s,z)=p(s,z)\bar f(s,z). \]

For h > 1:

\[ V_h(s,z)=p(s,z)\left[\bar f(s,z)+\gamma\sum_{z'}P(z'\mid z)V_{h-1}(s+1+\tau_z,z')\right] +(1-p(s,z))\gamma V_{h-1}(s+1,z). \]

The initial action is normalized by rounded relocation slots. The continuation policy waits in the reached zone; it does not optimize a new relocation action. This is truncated fixed-policy lookahead, not full finite-horizon Bellman optimality.

Corrected model-based MDP

The MDP module separately implements synchronous Bellman optimality backups. An action relocates from the current zone to a candidate zone. After arrival:

  • success earns fare, advances one pickup slot plus passenger duration, and moves to an OD-sampled dropoff zone;

  • failure earns zero, advances one pickup slot, and stays in the candidate zone.

Unlike the earlier implementation, every pickup attempt advances time and policy extraction depends on the current origin.

Static evaluation

The public answer for each query is a 263-zone reference-utility vector. Metrics are:

\[ NDCG@3=\frac{DCG@3}{IDCG@3} \]

and Hit@3, which checks whether the reference-optimal zone appears in the returned Top-3.

These metrics diagnose agreement with the reference objective. They do not measure observed driver behavior or causal revenue.

Rollout evaluation

The fixed simulator samples one executable recommendation with rank weights 0.6/0.3/0.1, relocates one driver, applies n/(n+40) pickup success in a concrete validation day/slot/zone cell, and samples a historical trip on success.

It omits competition, demand depletion, congestion, supply response, and equilibrium. Rollout values must be interpreted as within-simulator comparisons.

Statistical analysis

Strategies are compared with the same random seeds. Reports include:

  • mean and standard deviation;

  • paired bootstrap 95% confidence interval;

  • paired t-test;

  • Wilcoxon signed-rank test;

  • Cohen’s dz.

These quantify simulator Monte Carlo uncertainty only. Real-market inference requires day/week blocking and held-out temporal periods.

Counterfactual evaluation

IPS, SNIPS, and doubly robust estimators are implemented and tested. They are not applied to TLC trips because the data lack logged recommendation actions and behavior propensities. Reporting a numeric OPE value without those fields would be invalid.

Exposure analysis

Rank-weighted exposure is aggregated with the rollout weights. The report includes coverage, Gini, entropy/effective zone count, airport exposure, Manhattan exposure, and premium-fare-zone exposure.