Skip to content
Brocode SolutionsAI Software Development

FP&A · supply chain · credit risk

Forecast accuracy your CFO can act on.

SKU-store-week MAPE under 12%, PD / LGD models recalibrated to your current portfolio, or churn-at-risk lists your retention team can work tomorrow — built on your data, in your tenant, in twelve weeks.

MAPE

11.4%

−11.4pt vs SAS

Working capital days

38 → 29

−9.2 days

Forecast freshness

< 4h

overnight → near real-time

Forecast vs actual

Weekly demand · SKU-store-week

MAPE 11.4%
Historical Forecast (TimeGPT + LightGBM) Actuals to date

The problem at the next board meeting

27% MAPE locks AED 90M of working capital in slow-movers across 220 stores.

And the CFO has written to the Head of Supply Chain demanding a measurable improvement before the next quarterly board. For the credit-risk persona: the SME PD model is calibrated to a pre-COVID portfolio mix and is now systematically under-provisioning IFRS 9 stage-2 exposures.

An additional quarter of forecast inaccuracy locks another AED 25M into working capital and triggers a covenant-headroom conversation with treasury. The CFO has personally promised the board a working-capital improvement; missing it costs the FP&A lead executive credibility. Under-provisioned ECL is a CBUAE or SAMA inspection issue that lands directly on the Head of Credit Risk and the CRO. A 1.1-point ARPU drop in the customer base has analysts asking the CEO uncomfortable questions on the next earnings call.

A forecast is only useful if someone can act on it. We design predictive systems around the planning rhythm you already run — weekly stock reviews, monthly capital committee, quarterly board reporting — rather than imposed on top. That might mean an SKU-level demand forecast feeding directly into your ERP, a default probability score embedded in your loan origination workflow, or a maintenance risk index that schedules technician routes for the following week. Single-number forecasts give a false sense of certainty. We deliver calibrated confidence intervals, scenario overlays and backtesting evidence so reviewers can see how the model has performed against actuals over time.

Real forecasts in the region have to cope with Ramadan effects, school holidays, summer travel patterns, oil-price cycles and one-off government initiatives that rewrite demand overnight. Our teams build models that explicitly account for these factors rather than treating them as inconvenient outliers. We blend classical time-series methods with gradient-boosted trees and deep-learning architectures where they earn their keep, and we are equally happy explaining a SARIMAX model to a treasurer as we are tuning a Temporal Fusion Transformer for a retail planner.

The Forecast Tournament method

Six models. One holdout. Winner takes all.

Three weeks, fixed scope, free of charge. We run TimeGPT, Chronos, NeuralProphet, N-BEATSx, Temporal Fusion Transformer and LightGBM with engineered features against your last 24 months of data, then publish the result.

  1. Week 1

    Data ingestion + acceptance criteria

    Customer's last 24 months loaded into a sandbox in Snowflake or Databricks. Acceptance criteria signed: holdout strategy, MAPE definition, segments, the winner-takes-all rule.

    Charter signed

  2. Week 2

    Six models run in parallel

    TimeGPT, Chronos, NeuralProphet, N-BEATSx, Temporal Fusion Transformer and LightGBM with engineered features — same data, same holdout, same scoring. No favouritism, no parameter overfitting on the holdout.

    6 models, 1 holdout

  3. Week 3

    Winner-takes-all report

    Per-segment MAPE, per-segment ranking and a recommended production stack. The Brocode TimeGPT + LightGBM ensemble has hit 11.4% SKU-store-week MAPE against SAS Forecast Server at 22.8% on the shared dataset.

    Public methodology

  4. Weeks 4–10

    Production build

    Feature store on Feast, dbt transformations, MLflow tracking, drift detection on Evidently / WhyLabs. Real-time scoring via FastAPI + Triton where needed; batch scoring via Airflow on the customer warehouse.

    In-tenant deployment

  5. Weeks 11–12

    Parallel run vs the incumbent

    Parallel run against the customer's existing forecast (SAS, SAP IBP, Anaplan, o9, in-house). Stakeholders see the lift in their own terms before commit. Model card pack delivered with full SR-11/7 lineage.

    Stakeholder sign-off

  6. Run-phase

    Weekly accuracy review + scheduled recalibration

    Drift detector on every feature and the output. Quarterly recalibration cadence. Model cards updated against each release. Internal audit and regulator access on request.

    Quarterly recalibration

The model stack we deliver

Problem → recommended model family. No one-size-fits-all.

The right family per problem class, with the trade-offs we routinely explain to risk and finance reviewers.

ProblemRecommended familyNotes
Demand forecasting (SKU × store × week)TimeGPT + LightGBM ensemble · NeuralProphet baselineFoundation model for cold and short-history SKUs; LightGBM with engineered calendar / promotion / weather features for the tabular workhorse layer.
Inventory optimisationLightGBM demand + Gurobi / HiGHS stochastic optimiserService-level constraint exposed to the planner; output per location and SKU group.
Credit risk PDXGBoost with monotonic constraintsSHAP explanations sealed into model cards aligned to SR-11/7, BCBS 239, CBUAE / SAMA model-risk guidance.
Credit risk LGDSurvival analysis · DeepSurv + cox-PH baselineTime-to-default modelling with calibrated quantiles, ECL feeding straight into IFRS 9 stage-2 calculation.
ChurnLightGBM propensity + CausalML / EconML upliftTwo-stage so retention works prospects where treatment moves the needle, not high-propensity churners who would leave anyway.
Multivariate / hierarchicalTemporal Fusion Transformer + reconciliationFor multi-level forecasts (region × category × week) where coherence across the hierarchy is required.

Data and feature platform

In your tenant. On your warehouse. With the tools your team already runs.

Snowflake, Databricks or Microsoft Fabric for compute; Feast for the feature store; dbt for transformations; MLflow for tracking; Evidently and WhyLabs for drift. Great Expectations for data tests.

  • Snowflake
  • Databricks
  • Microsoft Fabric
  • BigQuery
  • dbt
  • Feast
  • MLflow
  • Evidently
  • WhyLabs
  • Great Expectations
  • FastAPI
  • Triton
  • Airflow
  • Power BI
  • Looker

Productionisation patterns

Real-time scoring on FastAPI + Triton Inference Server where latency demands it (pricing, risk-at-decision). Batch scoring via Airflow DAGs on the customer warehouse for daily / weekly cycles (demand, churn, ECL). Power BI or Looker dashboards for the FP&A or risk audience. Every prediction is logged with inputs, model version and confidence band, so auditors and risk reviewers can trace any decision back to evidence.

How we compare

SAS, SAP IBP / Blue Yonder / o9, offshore analytics, Big-4 advisory.

The structural differences and a published number from the shared tournament dataset.

CapabilityBrocodeSAS Forecast ServerSAP IBP / Blue Yonder / o9Offshore analytics shopBig-4 advisory
SKU-store-week MAPE on regional retail data

Shared 18,000-SKU / 320-store / 24-month tournament dataset.

11.4%22.8%No published number~15%~20%
Open-source ownership of the resulting model

Adapters, weights, configs and reproducer code all transfer to the customer.

Foundation forecasters (TimeGPT, Chronos) supported

Beyond ARIMA/ETS, into the 2024+ forecasting stack.

Regulator-grade model cards (SR-11/7, BCBS 239)

Aligned to CBUAE and SAMA model-risk guidance.

Standard delivery, every modelAdd-onAdd-onDIYStrategy pack
Co-existence with the planning suite (o9 / IBP / Blue Yonder)

We make the existing planner produce a better number.

Model partner, not a competing suiteReplacementN/AN/AStrategy pack
In-tenant deployment with no offshore data movement

UAE / KSA hyperscaler regions, customer-managed keys.

Build engagement vs advisory report

The Big-4 audit narrative is acknowledged; delivery is different.

Build — shipped models in 12 weeksBuild, but heavy licenceBuildBuild, offshore12-week assessment
60% run-cost reduction vs SAS estate

Removing SAS licence + moving to managed open-source pipelines on customer warehouse.

Three objections worth airing

SAS migration, MAPE in the long tail, and regulator acceptance on credit risk.

Objection 01

SAS has been our forecasting engine for 12 years. Why would we move to an open-source stack — and what is the migration cost?

You do not have to move SAS to see the lift. The Forecast Tournament runs on a sandbox of your data; the winning model is then put into a parallel run alongside SAS for one quarter before any cutover decision. The UAE retailer reference moved from SAS Forecast Server at 22.8% SKU-store-week MAPE to a TimeGPT + LightGBM ensemble at 11.4%, with an eleven-month payback and roughly a 60% reduction in annual forecasting run-cost. Migration is staged; SAS stays live until the open-source stack has earned its place in production.

Objection 02

We have had three vendors promise sub-15% MAPE. None held up at SKU-store-week granularity in long-tail SKUs. What is different here?

Two structural differences. First, foundation forecasters: TimeGPT and Chronos absorb cold-start and short-history SKUs that classical and tree-based models choke on. Second, the tournament: six models against the same holdout means we discover where each model wins and assemble the production ensemble per segment, rather than picking one model and forcing it across the catalogue. Long-tail performance is a published segment in the model card pack.

Objection 03

For credit risk, our model is regulator-approved. Any replacement must survive a CBUAE / SAMA model-validation review and have full lineage.

Model cards and lineage are standard delivery, not an add-on. XGBoost with monotonic constraints on PD survives SR-11/7, BCBS 239 and CBUAE / SAMA expectations. The KSA bank reference re-calibrated its SME PD model with Gini moving from 0.41 to 0.62; the ECL provision was rebalanced and accepted by the SAMA model validation team. We routinely walk through the validation pack with second-line risk before launch.

Free download

The GCC Demand-Forecast Tournament — 6 Models, 18,000 SKUs, 320 Stores, 24 Months

A 36-page PDF, the per-model scoreboard, and a downloadable model-card pack (PDFs and JSON) for each contender. Headline: TimeGPT + LightGBM ensemble at 11.4% SKU-store-week MAPE versus SAS Forecast Server at 22.8% on the shared dataset, with 4× faster batch scoring.

  • Tournament methodology — holdout strategy, scoring, acceptance criteria
  • TimeGPT, Chronos, NeuralProphet, N-BEATSx, TFT, LightGBM head-to-head
  • Long-tail SKU performance segment-by-segment
  • Model card pack aligned to SR-11/7 and BCBS 239
  • Working-capital impact on the retailer reference (−9.2 days)
  • Move-to-production path with parallel-run rules

Instant download. No spam. Unsubscribe any time.

Frequently asked

Data, regulator, seasonality, drift, residency.

Eight questions a CFO, a Head of Credit Risk and a CISO usually share with internal audit before procurement.

  • 24 months of clean transactional data is the comfortable floor for SKU-store-week demand forecasting; 36 months covers the regional calendar effects in full. For credit risk PD, the cleaner number is 5,000+ defaults across the cycle. For churn, 12 months of subscriber events. We can run a feasibility pass on smaller histories before committing.

Forecast Tournament

Three weeks, fixed scope, free of charge.

Six models against your last 24 months of data. Winner takes all on MAPE at your chosen granularity. Decision to move to production stays with you.

What you receive

  • · Per-segment MAPE for six contender models
  • · Model card pack for each contender (SR-11/7 aligned)
  • · A recommended production stack with cost envelope
  • · A 45-minute review with the forecasting lead

Quote request

Run a free 3-week Forecast Tournament on your data

We take your last 24 months, run six forecasting approaches, and publish the results. Winner picked on MAPE at your chosen granularity. No commitment until you see the numbers.

Prefer chat? Message us on WhatsApp — we'll see it within working hours.

Forecast TournamentWhatsApp