This research proposes a novel demand forecasting method which will work effectively even in such circumstances where extrapolate-able demand patterns are hardly available. The method uses the market mechanism to aggregate tacit knowledge of the firm's sales people on the future demand of a product into a continuous forecasted demand distribution. A new type of prediction security and an original market maker algorithm suitable for the security are introduced and furnished into an intra-firm prediction market system. As a result, sufficient liquidity is secured for the market even when the number of the traders is small, and the market maker can output at any time an aggregated demand forecast of the traders as a continuous distribution. An agent simulation model, where each trader has learning capability, is also developed to study how the method works, whether the output distribution is adequate, how quickly the output converges, etc.