Predictive LTV

Predictive LTV is an analytical/ML project (with refresh): based on customer history and early behavior we train a model that estimates the expected future value of a customer or segment, and link it to CAC to assess payback. Unlike historical LTV (what a customer has already brought), here we forecast the future from early signals. Honestly about the nature, this is key: this is a PROBABILISTIC estimate with confidence intervals, NOT an exact prediction and NOT a guarantee — every forecast has a margin of error, individual customers may behave differently, and the model is most accurate 'on average for a group', not for a specific person. Honestly about data: sufficient history with completed lifecycles is needed for training — on short history or small numbers the model is unreliable; dirty data = a wrong forecast (garbage in, garbage out). Honestly about aging: behavior and the market change, the model degrades — it needs retraining and forecast-vs-fact reconciliation. Honestly about the essence: this is an estimate for decisions, NOT raising LTV and NOT a revenue guarantee — the model helps decide whom to invest in and how much to pay for acquisition, but raising LTV (product, retention, upsell) is done by you; the forecast itself does not increase value. Honestly about metrics: the model relies on correlations, not proven causes. Requirements: access to data and agreed metrics. An important boundary: this is predictive LTV, not historical LTV/CLV modeling (676 — separate/simpler), not a churn model (separate) and not retention work. If data is scarce or the business is new, predictive LTV is premature. Picture this: instead of 'customer value is clear only after a year' you assess promise from early behavior already — as an estimate with a range. The base price starts from 30,000 ₽ per project; it depends on data and model complexity.

• Customer value is visible only after the fact, months later.
• You pay for acquisition without estimating future value.
• You cannot tell promising customers from one-off ones early.
• You make investment-in-customers decisions without a forecast.

• Preparing history and early-behavior features
• Training a predictive LTV model
• Confidence intervals and accuracy estimate
• Forecast by segments/channels
• Linking to CAC and payback assessment
• A report with assumptions, ranges and limitations
• Reviewing results with you
• Model-refresh recommendations

• An estimate of future customer value from early behavior (with a range)
• You understand whom it is reasonable to invest in
• An acquisition-price ceiling by promise
• A base for decisions (no LTV growth or guarantee)

• We clarify the goal, data, metrics, horizon; collect access
• We prepare data, train and check the model, compute intervals
• We compile a report with ranges and limitations, review with you

• Will the LTV forecast come true for each customer?

  No. It is a probabilistic estimate with a margin of error; it is most accurate 'on average for a group', while an individual customer may behave differently. We give a range and accuracy estimate, not a per-person guarantee.

• Will the model raise LTV?

  No. It estimates future value and helps decide whom to invest in, but by itself it does not raise LTV. Growth is product, retention, upsell (separate work); the forecast only guides decisions.

• How is this different from historical LTV?

  Historical modeling counts what a customer has already brought; predictive LTV forecasts future value from early signals. The forecast is more complex and needs more data; on short history it is unreliable.