A five-model ML ensemble forecasts every account fourteen days ahead, raises alerts before deficits happen, and explains each prediction with gpt-4o-mini.
A fintech treasury team manages dozens of accounts across currencies and banks. Clearing delays cause cash gaps. To stay safe they freeze millions in idle reserves. They notice problems only after they happen, then scramble to rebalance.
Sight gives them a forecast, a reason, and a fix — for every account, every day, in real time. Four steps below.
Drag Time Machine. Every account balance, every status colour, every alert recomputes from real ML predictions fourteen days ahead. USD-NYC is fine today, $150K below minimum on day five — you watch it happen.
Auto-generated alerts the moment a forecast dips below an account minimum. Severity escalates with how low the dip goes. No manual monitoring of eleven accounts across five currencies.
Click any alert, get a 2-sentence explanation from gpt-4o-mini using the actual forecast numbers. Why the dip is happening, when it bottoms out, and what to do about it.
Sight Compass finds the optimal donor account, picks SEPA or SWIFT based on currency pair, sizes the transfer to cover the deficit plus a 20% buffer, executes it. Balance restored, forecast cleared.
Every major treasury blow-up of the last decade — SVB, Credit Suisse, Synapse, the London Whale — shared the same pattern. A risk was visible in the data. Nobody had the tools to see it in time.
Sight is built directly against this pattern. Each feature below is a response to a documented failure, grounded in the same industry research that treasury benchmarking firms publish every year.
Overnight illiquidity. Many fintechs held 80%+ of cash in a single counterparty.
Concentration Card (HHI) — tracks share of portfolio in any one bank.
Banking-as-a-service ledger discrepancies. Internal records desynced from real correspondent balances.
Versioned JSON contract between ML and UI. Conformal P10/P90 instead of point estimates.
Counterparty risk materialized worldwide. Treasurers had hours to reposition.
Counterparty dimension in HHI. Roadmap: contagion graph for second-order exposures.
Excel copy-paste error in VaR model. Risk math used sum instead of average — passed review.
Liquidity Gradient solver in typed code. No Excel. All math is tested and versioned.
Batch inference — the same architecture Spotify, Netflix and Bloomberg use for prediction at scale.
365 days of NovaPay transactions across 11 accounts. Clearing delays, holidays, weekly cycles — all encoded.
Five ML models train in parallel: Prophet, LightGBM with quantile loss, ARIMA, Holt-Winters ETS, Amazon Chronos.
Stacked Ridge meta-learner picks optimal weights per account. Output: a static forecasts.json with 14-day predictions.
Next.js reads the JSON, renders the 3D globe, Time Machine slider, alerts, and AI insights — no backend, no Python at runtime.
Each model is wrong in its own way. Per-account selection picks the best candidate from honest OOF holdout — Chronos wins 9/11, Ridge stacker 1, Prophet 1.
Trend, weekly seasonality, country holidays
Cross-account features, lag interactions, payday cycles
Short-term autocorrelation
Exponential smoothing with trend + seasonal
Foundation model, patterns from millions of time series
Ridge stacker is trained on rolling-origin out-of-fold predictions (5 windows × 7 days = 35 OOF pairs per account) and evaluated on a separate held-out test window. For each account we then pick the lowest-MAPE candidate — base model or stacker.
Thirty walk-forward cutoffs over 365 days of history. Trained on data before each cutoff. Predicted what happened after. Compared against three baseline forecasters. Every digit on this page is reproducible from public/data/backtest_results.json.
Customer development
Composite quotes synthesized from CustDev conversations and secondary research (PwC, Deloitte). Anonymized by design — each persona maps to a real role, a real corridor, and a specific Sight feature.
“Reconciling nostro balances and SWIFT confirmations in Excel takes my team two business days a week. When our USD balance in New York drops below the floor, we learn about it from a bank email — that's six to eight hours of scramble and sometimes regulator-facing penalties.”
Two AI layers stacked. Predictive answers “what will happen”. Generative answers “why and what to do”.
The account balance is forecasted to drop to $150,233 by May 23, which is significantly below the required minimum of $800,000. This stems from cumulative low forecasted balances on May 22 ($244K) and May 23 ($150K). To mitigate, consider increasing liquidity through funding or reducing outflows before these dates.
A liquidity-gradient algorithm routes funds from low-pressure surplus accounts to high-pressure deficit accounts, in the right currency, through the cheapest channel.
