Research Works Supporting Our Solution

This paper helped us anchor our entire literature survey and justify our core design choice of using machine learning over traditional statistical methods for forecasting highly volatile freight rates. From this work we take the structured comparison of ML approaches across the maritime value chain, which we use to shortlist candidate models (time series, regression, and deep learning families) for our forecasting engine and to defend that choice in front of the jury.

This is the closest published architecture to what we are building, so we take its ensemble deep learning design that forecasts multiple freight indices at once while injecting external factors like economic indicators and uncertainty. We adopt its interpretability layer so that our dashboard does not just output a predicted rate but also explains which external drivers moved the forecast — exactly what a logistics manager needs before committing to a short-term or mid-term charter.

From this paper we take the hybrid "model pool" idea: maintain a pool of time series, traditional nonlinear, and deep learning models, and use error evaluation with reinforcement learning to pick the best performer per route and per vessel class. This directly shapes our solution segment where a single forecasting model would fail on volatile routes — our system dynamically routes each prediction to the strongest model, improving accuracy without hand-tuning.

This open-access paper gave us a reproducible RNN–GRU baseline for forecasting a regional freight index, which we adapted as our quick prototype for route-specific lanes such as Australia to Paradip or Indonesia to Haldia. We take its demonstration that a comparatively simple GRU network can already capture regional freight dynamics, which lets us ship a working minimum viable model during the hackathon and layer the ensemble and constraint logic on top of it.

This paper helped us answer the jury question "why not just use ARIMA or econometric models?" by providing a head-to-head comparison of classical time-series econometrics against ML/AI frameworks. From it we take the benchmarking discipline itself — our solution reports both an econometric baseline and ML forecasts side by side so the improvement is measurable and credible, and we borrow its feature set of key shipping-sector drivers as candidate inputs to our model.

This paper supports our real-time port congestion module by showing how deep learning can predict congestion from an international port network perspective rather than port by port. We take its network-level view to build the risk early-warning feature of our dashboard, so congestion building up at an origin or discharge port raises a flag before it distorts our chartering recommendation — directly serving the risk mitigation requirement of the problem statement.

This highly cited work gave us the concrete technique for estimating berth time and port time from AIS data, which is the backbone of our turnaround-time estimation for East Coast Indian ports like Paradip, Vizag, Gangavaram, Dhamra and Haldia. We take its congestion-status-to-time-in-port pipeline so our vessel type optimizer can reject a large Capesize not only on draft grounds but also on predicted waiting time, minimizing idle time as the problem statement demands.

From this open-access paper we take a practical, deployable machine learning approach to predicting vessel waiting times, which feeds our idle scenario management segment. When the model forecasts a low-demand or high-congestion window, our system uses these waiting-time estimates to suggest alternative employment, repositioning, or delayed market entry — turning the "idle scenario management" requirement from a static rule table into a data-driven recommendation.

This paper is the bridge between our two hardest modules: it proves that port congestion data materially improves freight rate forecasts and that the resulting predictions can inform chartering decisions. We take its RBF neural network coupling of congestion indicators with freight indices so our forecasting engine ingests live congestion as an exogenous variable — meaning a congestion spike at an origin port automatically shifts the predicted rate and the recommended market entry window.

From this paper we take the insight that congestion does not stay local — it propagates across the port network — and we bake that into our disruption early-warning logic. If congestion is building at a major transshipment hub or loading origin like Australia or Indonesia, our system treats it as a leading indicator for delays and rate movement on the East Coast India lanes, giving chartering managers advance notice instead of same-day bad news.

This is our primary reference for the risk mitigation deliverable, because it applies machine learning to forecast volatility itself — not just the rate — and benchmarks it against classical time-series models across different shipping markets. We take its volatility-forecasting layer so our dashboard can warn "the Handysize segment on this lane is entering a high-volatility regime" and recommend locking a mid-term contract or waiting, which is precisely the proactive, predictive chartering strategy the problem statement asks for.

This paper gave us the regime-switching concept that anchors our optimal market entry timing feature: freight markets alternate between low-cost and persistently high-cost regimes, and the trick is to enter contracts before the regime flips. We take its regime detection approach to classify the current market state for each route and vessel class, so our recommendation engine can say "enter a short-term contract now, regime shift likely in 3–5 weeks" instead of reacting to daily market noise.

Since our cargo is bulk commodities like coal and iron ore, this paper supports the commodity price trend segment of our feature engineering. We take its hybrid deep-learning-plus-optimization approach for commodity price prediction and its explicit linkage of commodity indices to Baltic Dry Index shipping costs, so our model uses commodity price movement as a leading signal for freight rate direction on each origin-destination pair.

This work speaks directly to our stated objective of moving from multiple single spot contracts to short-term and medium-term multiple voyage contracts. We take its comparative framework of time charter versus spot strategy across vessel types — Handysize, Supramax, Panamax and Capesize — to structure our recommendation logic: when the forecast shows a rising rate regime, lock a time charter; when the forecast shows softness, stay spot for a single voyage.

Although set in offshore operations, this paper contributes the optimization-under-uncertainty machinery we adapt for chartering cost minimization: modeling uncertain conditions spatially and choosing chartering windows that minimize total expected cost. We take its stochastic planning formulation as the mathematical skeleton behind our "optimal market entry timing" recommendation, replacing weather uncertainty with freight rate and congestion uncertainty.

This paper grounds our solution in the dry bulk reality of the problem statement — coal and other bulk cargo moved by Handysize through Capesize vessels — and shows how digitalization and AI optimize bulk carrier operations and management. We take its framing of digitalization as an efficiency lever to justify our dashboard-centric design, where port constraint data, vessel specs and forecasts live in one interface rather than in daily phone calls to the market.

This study of shipping cycle patterns and their forecast ability informs the long-horizon layer of our model, distinguishing cyclical movement from short-term noise. We take its cycle-aware perspective so that when our system recommends a medium-term multiple voyage contract, the recommendation is backed not only by next-week rate forecasts but also by where the analysis says we are in the broader shipping cycle for bulk tonnage.