Archived final year project, kept public deliberately. This was my dissertation project for BSc Computer Science at the University of Westminster. The code runs and is reasonably clean, but both models contain methodology errors I did not catch at the time. I have left them unfixed and documented them in full below, because the post-mortem is more useful than the results were.
Do not read the reported metrics as working results. See What I'd do differently.
Notebook: BTC_Classification.ipynb
Random Forest classifier over 5-minute BTCUSD bars.
- Data:
BTCUSD_m5.csv— Jan 2023 to May 2024, ~110,000 rows, sourced from Dukascopy - Task: Three-class trend label — no trend (0), downtrend (1), uptrend (2)
- Features: ATR, RSI, midprice, SMA 40/80/160, and rolling linear-regression slopes of each over a 6-bar window
- Split: Chronological 70 / 15 / 15
- Reported result: 75.3% validation accuracy, 71.5% test accuracy
Notebook: LSTM_BTC_Final_Model.ipynb
Two-layer LSTM regression on close price. The effective forecast horizon is two bars, not one — see error 5.
- Data: 1-minute BTC-USD bars pulled live via
yfinance(rolling 7-day window) - Features: OHLC, RSI(15), EMA 13/50/200
- Architecture: LSTM(32, return_sequences) → Dropout(0.2) → LSTM(32) → Dropout(0.2) → Dense(1); Adam, MSE loss, early stopping on validation loss
- Sequence length: 4
- Split: Chronological 80 / 10 / 10
- Reported result: validation R² 0.987, test R² 0.013
Reviewed in 2026. Five substantive problems, in order of severity.
check_candles labels row i by comparing Close[i-6:i] against
MA40[i-6:i] — entirely past bars. The model classifies a trend that has
already happened; it does not forecast anything. Grouping the labels by
subsequent return makes this plain:
| Label | Mean past 6-bar return | Mean future 6-bar return |
|---|---|---|
| Uptrend | +0.0916% | +0.0134% |
| Downtrend | −0.0740% | +0.0054% |
| No trend | −0.0021% | +0.0032% |
The past separates cleanly. The future is near-identical across all three classes — the label carries almost no information about what happens next. The 71.5% is real accuracy at recognising a label derivable from data the model can already see.
Close > MA40 — one comparison, no model, no training — scores 67.5% on
the same test split against the Random Forest's 71.5%. Fifteen features and
an ensemble bought four points. The majority-class floor is 37.2%, which is
the comparison the notebook implicitly invites and which flatters the result
badly.
The MinMax scaler is fitted on the full dataset before splitting:
df_scaled = scaler.fit_transform(df[features + ['TargetNextClose']]) # all rowsThe symptom is visible in the committed output — Test MAPE: inf. That only
occurs when y_test contains exactly 0.0, which in MinMax space means the
global minimum of the entire series fell inside the test window and the
scaler had already seen it.
Test R² of 0.013, against a persistence baseline (next close = current close, zero parameters) scoring R² 0.997, MAE ~$84. The model is dramatically worse than no model.
The validation R² 0.987 → test R² 0.013 collapse was the warning sign, and I did not act on it. Metrics were also never inverse-transformed, so the reported RMSE of 0.0748 is in scaled 0–1 units and means nothing in dollars.
create_sequences builds each sample from rows i to i+3 and takes its
target from TargetNextClose at row i+4. Since TargetNextClose is
Close.shift(-1), the target is Close[i+5] — two bars beyond the last
input bar, with bar i+4 used neither as an input nor as the target.
The variable name describes what I intended, not what the code does. Unlike the errors above, this one made the task harder rather than flattering the results: a two-bar horizon is strictly more difficult than one, and the model was denied the intervening bar entirely. It does not explain the poor test performance — the scaler leak does — but it means the model was never evaluated on the problem the report claims it solved.
- Non-stationary features into a tree. Raw OHLC and moving-average
levels fed to a Random Forest trained on $16–30k prices and tested on
$57k+. Trees split on absolute thresholds and cannot extrapolate beyond the
training range. Ratios (
Close/MA40 - 1) would have been correct. - Unexamined data quality. 3.1% of bars have zero price change, with a longest frozen run of 252 consecutive bars — about 21 hours of flat synthetic price from Dukascopy gap-filling. Never cleaned or acknowledged, and it inflates the "no trend" class.
- Irreproducible by construction. The LSTM pulls a rolling 7-day
yfinancewindow, so it returns different data on every run. Combined withrandom_state=Noneon the classifier and no TensorFlow seeds, none of the reported numbers can be regenerated — including by me. StandardScalerbefore a Random Forest is a no-op. Harmless, but it shows the pipeline was assembled by pattern rather than by reasoning.Volumedropped from the LSTM features with no stated justification.
Errors 1 to 4 are one mistake wearing four hats: not printing a trivial baseline next to the model. A backward-looking target and a leaking scaler both look like success right up until something with zero parameters beats you. A baseline is now a precondition for reporting any result, not a nice-to-have.
Error 5 is a different failure and worth separating out. Nothing about the
output looked wrong, because the bug was an off-by-one between a variable's
name and its contents. TargetNextClose was taken at its word — by me then,
and by every subsequent reading of the notebook — until someone traced the
indices by hand. Naming is not documentation, and a pipeline that transforms
indices needs its alignment asserted in code, not inferred from a label.
Note also that errors 1 to 4 all flatter the results while error 5 works against them. They do not all point the same way.
The notebooks are archived as-is and are not reproducible — see point 4 above. If you want to run them anyway:
git clone https://github.com/hmzeda6210/Crypto_Predictions.git
cd Crypto_Predictions
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
jupyter notebookTA-Lib requires a system-level C library that pip will not install for
you. On macOS: brew install ta-lib. On Debian/Ubuntu you need to build it
from source. On Windows, use a prebuilt wheel. The LSTM notebook imports it
but barely uses it — if installation is painful, you can drop the import talib line and the notebook still runs.
BTC_Classification.ipynb runs offline against the committed CSV.
LSTM_BTC_Final_Model.ipynb requires an internet connection and will fetch
whatever the last 7 days happen to be.
| File | Description |
|---|---|
BTC_Classification.ipynb |
Random Forest trend classification |
LSTM_BTC_Final_Model.ipynb |
LSTM close-price regression |
BTCUSD_m5.csv |
5-minute BTCUSD OHLC, Jan 2023 – May 2024 (Dukascopy) |
RandomForest_Pipeline_Diagram.png |
Classification pipeline diagram |
LSTM_Pipeline_Diagram.png |
LSTM pipeline diagram |
MIT — see LICENSE.