How to Do Time Series Forecasting Without Code

TensorTurn forecasts any time series by turning it into a supervised learning problem: you upload a timestamped metric, and it engineers lag, rolling and calendar features, then trains a regression model (XGBoost or scikit-learn) or a sequence network (PyTorch or Keras), and deploys a /predict API. To be clear and honest, this is a machine learning approach rather than a built-in ARIMA or Prophet statistical toolkit; it excels when there are drivers, nonlinear patterns and enough history, and it needs time-based validation to avoid leakage.

How TensorTurn turns a time series into a model

Rather than fitting a classical statistical model, TensorTurn reframes forecasting as supervised regression. Each timestamp becomes a row, and the past becomes features: lagged values (last period, same period last year), rolling averages and standard deviations, and calendar terms (day of week, month, holiday). A model then learns how those features map to the next value. This works for one metric (univariate) or many drivers at once (multivariate), and it naturally captures nonlinear effects that linear statistical models miss.

What data you need

• A timestamp column at a regular interval (hourly, daily, weekly, monthly) plus the target metric to forecast.
• At least two full seasonal cycles of history so seasonality can be learned.
• Optional exogenous variables (drivers) that influence the target — but you must know their future values to use them at prediction time.
• Notes on gaps, holidays and regime changes so they can be handled rather than learned as noise.

How to build it on TensorTurn (no code)

• Import your time-stamped data (CSV, Excel or from a URL).
• Run the health check for outliers, mixed types and gaps.
• In chat, state the target, cadence and horizon: 'Forecast daily active users 14 days ahead.'
• TensorTurn engineers lag, rolling and calendar features and trains the model, self-healing any errors.
• Evaluate with a time-based backtest — never a random split — reviewing MAE, RMSE and MAPE plus a forecast-vs-actual plot.
• Deploy a /predict endpoint and feed it new periods to produce forecasts.

Validate on time, not at random

The most common mistake in forecasting is shuffling rows into a random train/test split, which lets the model peek at the future and reports fantasy accuracy. TensorTurn evaluates with a time-based split so the test period genuinely comes after the training period, mirroring how the model will be used. If a result looks too good to be true, check that the validation respected time order.

How accurate can it be?

For stable series with clear seasonality, a MAPE of 5–15% is realistic. For volatile or event-driven series, expect 20–40% or more. Two rules always hold: accuracy degrades as the horizon grows (one step ahead is far easier than thirty), and no model can forecast genuine shocks like a sudden outage or a viral spike. Treat short horizons as reliable and long horizons as directional.

Deploy as an API

Publish the trained forecaster as an authenticated /predict endpoint on Modal that scales to zero when idle, secured with a SHA-256-hashed Bearer key and per-call logging. The playground provides cURL, JavaScript, Python and Rust snippets so you can pull forecasts into dashboards or planning systems.