Titane Intelligence is releasing TITAN-HORIZON 2.5, the new generation of its sales prediction engine for e-commerce sites. This version doubles the forecast horizon (12 weeks instead of 6, while keeping the same accuracy), and adds six capabilities absent from the previous generation — including a scenario simulator (marketing budget, email campaigns, promotions, product launches), the forecasting from launch of new products with no history, and per-product reasoning that spells out, for each prediction, the real drivers behind the forecast.
What TITAN-HORIZON is for
TITAN-HORIZON is Titane's demand forecasting engine. For every product in the catalog, the model predicts sales in units and revenue in euros over the coming weeks, with a calibrated confidence interval.
Where the model's accuracy comes from: TITAN-HORIZON cross-references 28 signals per product — sales history and stock levels, but also live advertising campaigns (Meta, Google), 14-day weather, the commercial calendar (Sales, Black Friday, holidays), and weak market signals (Google Trends, competitor stockouts). The model sees what an Excel file cannot see.
The forecasts activate everywhere a decision is made on the basis of a future volume: steering supplier orders, sizing safety stock, validating marketing budgets, forecasting quarterly revenue, alerting on stockout and overstock risks. The model relies on a proprietary deep learning architecture calibrated specifically on each site's data.
Overview — what changes with 2.5
The table below situates TITAN-HORIZON 2.5 relative to the market's classic approaches (moving averages, ARIMA, business heuristics in Excel) and relative to the model's two previous generations.
Migrating from the previous generation
For clients already equipped with a previous generation, 2.5 plugs into the same connectors and the same data sources: no overhaul on the site, ERP or steering tool side. The input data stays identical (sales history, product catalog, exogenous features such as advertising, weather, commercial calendar).
Three operational changes are worth noting. The output format gains a forecast_raisonnement field that lists the explanatory factors of the prediction, per product, and a scenarios_marketing module that holds the custom-built simulations. Finally, the confidence intervals provided are now calibrated and audited : the real accuracy (91%) is measured at each deployment, and published in the performance report delivered to the client.
The six key advances
Six capabilities have been added compared to the previous generation. Each one answers a need observed in production on the first TITAN-HORIZON deployments.
Before / After: what it changes in the day-to-day conversation
Before — “The model forecasts €1.2M this quarter.” — “And for what reasons? What are the factors?” — “That's a good question.”
With TITAN-HORIZON 2.5 — “The model forecasts €1.2M this quarter, within a range of €1.05M to €1.35M at 80% reliability. This projection rests 60% on the June Meta campaign and 25% on the summer Sales effect. If we cut the Meta budget by 25%, we drop back to €1.05M. If we double it, we climb to €1.5M but we take on a stockout risk on 12 products before the restock. Here's the list.”
That's a different conversation.
Where it activates in practice
TITAN-HORIZON 2.5 produces, for every product and every week, a quantified, explained forecast accompanied by a recommended action. The outputs are delivered in La Forge Titane and/or in the site's existing software. The model is deployable everywhere — here are the most widely deployed activation areas:
Supply & Purchasing steering. Weekly list of products to restock, sorted by urgency, with order deadline and recommended quantity. Buyers work from a prioritized queue, not an Excel file to sort by hand.
Marketing budget validation. Before launching a Meta or Google campaign, the "What-if" module produces the projection of sales, revenue and stockout risks for each product.
P&L steering and reporting. 12-week revenue forecast with confidence interval, broken down by brand and by category.
New-product launches. Forecast available from the very first week online, with an honest confidence interval (re-simulation of 1,000 trajectories). The launch stock is sized on a statistical basis, not on a bet.
How it's evaluated
TITAN-HORIZON 2.5 is evaluated on the sales actually observed after the forecasts are generated, with no data leakage (strict temporal split). The metrics tracked are:
Two new proofs of value are integrated in 2.5. Comparison against naive baselines : the model is systematically benchmarked against three simple methods (8-week average, 16-week average, same week last year). It only goes to production if it beats them significantly on accuracy. Multi-window temporal backtest : the model is tested on three distinct historical periods to verify that it is robust over time, not just performant on one window by luck.
How it's deployed
The model runs automatically on cloud infrastructure managed by Titane. The forecasts are delivered in La Forge Titane, or directly in the client software. The same output feeds the marketing tools (Mailchimp, Klaviyo, Brevo) for elasticity calculations and campaign planning.
No migration. No overhaul of the existing setup. Titane handles the entire process: data ingestion, feature engineering, training, deployment, monitoring.
First production deployment
TITAN-HORIZON 2.5 has been deployed with a first Titane client on a catalog of several hundred SKUs. The confidence interval coverage measured at 91%, and the accuracy held over the 12-week horizon, are the two calibration results audited immediately upon delivery. The first measured business results — gains from avoided stockouts, margin gains from inventory steering, marketing gains from budget allocation — will be the subject of an upcoming deployment article.
What TITAN-HORIZON 2.5 does not do yet
Three directions remain ahead in the coming versions and are already the subject of early internal tests.
Process automation
Today TITAN-HORIZON delivers order recommendations. The next step: writing back into the client software, triggering the order directly from the ERP, closed-loop, under human validation.
Multimodality
The model will understand the visuals used on the website, on product pages, in email campaigns and in Ads. These iconographic signals will become a forecast input, to gain precision on visually-driven products and high creative-intensity periods.
Additional signals
Several additional signal sources are being implemented to gain even more forecast precision, without weighing down the integration on the client side.
What's next
The main effort of the coming weeks is on the deployment of TITAN-HORIZON 2.5 across our clients : measuring the first business results on avoided stockouts and marketing gains, then multiplying the activation surfaces on each site (Supply, Marketing, Finance, Product) to put the model to work across the entire commercial cycle.
