Abstract:
Problem: Cryptocurrency markets are prone to abrupt trend reversals, yet most predictive
models still focus on price direction or volatility rather than the detection of market bottoms—
the earliest, tradable inflection points that mark the shift from decline to recovery. This leaves
traders without timely signals to re-enter the market, amplifying the financial and psychological
toll of crashes and limiting the effectiveness of automated trading strategies.
Methodology: Our study proposes search-informed transformer based crypto bottom predictor,
a two-stage pipeline that first learns realistic future price paths and then predicts whether the
very next bar is a market bottom. In the first stage, a Temporal Fusion Transformer acts as the
price predictor. It ingests multivariate inputs—OHLCV history, engineered technical
indicators—to produce multi-horizon price forecasts. The predictor’s final hidden state, a dense
summary of latent market forces, becomes the chief feature set for stage 2. Here, an XGBoost
classifier combines those latent vectors with the most recent Google-Trends scores to output
the probability that the next time-step is a bottom. The design therefore links generative
forecasting with explicit bottom detection, enabling an early-warning signal.
