megaproxy/MyMidas
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ML predictions Phase 4: SARIMA spending forecast with dual confidence bands

Browse source 
Replaces unused Prophet dependency (unrunnable without cmdstan) with
SARIMA (statsmodels SARIMAX) as the primary spending forecast algorithm.
Strategy: SARIMA(1,1,1)(1,0,1,12) for 12+ months of data, ARIMA(1,1,1)
for 6-11 months, Holt-Winters for 3-5 months, simple average below that.

Adds 95% confidence bands (1.96σ) alongside existing 80% (1.28σ).
Extends forecast horizon from 3 to 6 months and actuals display from
6 to 12 months. Each category now carries an algorithm field surfaced
as a badge in the UI. Frontend chart shows both confidence tiers as
stacked bar overlays with a 3-month summary grid below.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
megaproxy 2026-04-28 10:30:26 +00:00
parent 3b4787d8b9
commit 4572621f5d
4 changed files with 109 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

91
backend/app/ml/spending_forecast.py
View file

@ -10,19 +10,55 @@ import pandas as pd
warnings.filterwarnings("ignore")
MIN_POINTS = 3
FORECAST_MONTHS = 3
FORECAST_MONTHS = 6
def _next_month_starts(from_date: date, n: int) -> list[str]:
months = []
d = (from_date.replace(day=1) + relativedelta(months=1))
d = from_date.replace(day=1) + relativedelta(months=1)
for _ in range(n):
months.append(d.strftime("%Y-%m-%d"))
d += relativedelta(months=1)
return months
def _fit_holt(values: list[float], n: int) -> tuple[list[float], list[float], list[float]]:
def _fit_sarima(values: list[float], n: int) -> tuple[list[float], list[float], list[float], list[float], list[float], str]:
"""
Primary algorithm. Uses SARIMAX with seasonal component when enough data exists,
plain ARIMA otherwise. Returns (forecast, lower_80, upper_80, lower_95, upper_95, algorithm).
"""
from statsmodels.tsa.statespace.sarimax import SARIMAX
series = np.array(values, dtype=float)
algo = "sarima"
try:
if len(series) >= 12:
# Seasonal ARIMA with annual period
model = SARIMAX(series, order=(1, 1, 1), seasonal_order=(1, 0, 1, 12),
enforce_stationarity=False, enforce_invertibility=False)
else:
model = SARIMAX(series, order=(1, 1, 1),
enforce_stationarity=False, enforce_invertibility=False)
fit = model.fit(disp=False, maxiter=200)
forecast_obj = fit.get_forecast(steps=n)
mean = forecast_obj.predicted_mean
ci_80 = forecast_obj.conf_int(alpha=0.20) # 80% interval
ci_95 = forecast_obj.conf_int(alpha=0.05) # 95% interval
lower_80 = np.maximum(0, ci_80.iloc[:, 0].values).tolist()
upper_80 = ci_80.iloc[:, 1].values.tolist()
lower_95 = np.maximum(0, ci_95.iloc[:, 0].values).tolist()
upper_95 = ci_95.iloc[:, 1].values.tolist()
return mean.tolist(), lower_80, upper_80, lower_95, upper_95, algo
except Exception:
return _fit_holt(values, n)
def _fit_holt(values: list[float], n: int) -> tuple[list[float], list[float], list[float], list[float], list[float], str]:
"""Holt-Winters fallback."""
from statsmodels.tsa.holtwinters import ExponentialSmoothing
try:
@ -36,13 +72,22 @@ def _fit_holt(values: list[float], n: int) -> tuple[list[float], list[float], li
fit = model.fit(optimized=True, disp=False)
forecast = fit.forecast(n)
sigma = float(np.std(fit.resid)) if len(fit.resid) > 1 else float(np.mean(values) * 0.15)
lower = np.maximum(0, forecast - 1.28 * sigma)
upper = forecast + 1.28 * sigma
return forecast.tolist(), lower.tolist(), upper.tolist()
lower_80 = np.maximum(0, forecast - 1.28 * sigma).tolist()
upper_80 = (forecast + 1.28 * sigma).tolist()
lower_95 = np.maximum(0, forecast - 1.96 * sigma).tolist()
upper_95 = (forecast + 1.96 * sigma).tolist()
return forecast.tolist(), lower_80, upper_80, lower_95, upper_95, "holt_winters"
except Exception:
avg = float(np.mean(values))
sigma = float(np.std(values)) if len(values) > 1 else avg * 0.15
return [avg] * n, [max(0, avg - 1.28 * sigma)] * n, [(avg + 1.28 * sigma)] * n
fcast = [avg] * n
lower_80 = [max(0.0, avg - 1.28 * sigma)] * n
upper_80 = [(avg + 1.28 * sigma)] * n
lower_95 = [max(0.0, avg - 1.96 * sigma)] * n
upper_95 = [(avg + 1.96 * sigma)] * n
return fcast, lower_80, upper_80, lower_95, upper_95, "average"
def forecast_spending(df: pd.DataFrame) -> list[dict]:
@ -61,31 +106,47 @@ def forecast_spending(df: pd.DataFrame) -> list[dict]:
group = group.sort_values("ds")
values = group["y"].tolist()
actuals = [
{"date": row["ds"].strftime("%Y-%m-%d"), "amount": row["y"]}
{"date": row["ds"].strftime("%Y-%m-%d"), "amount": round(float(row["y"]), 2)}
for _, row in group.iterrows()
]
if len(values) < MIN_POINTS:
avg = float(np.mean(values))
sigma = avg * 0.15
forecast_pts = [
{"date": d, "amount": round(avg, 2), "lower": round(avg * 0.7, 2), "upper": round(avg * 1.3, 2)}
{
"date": d,
"amount": round(avg, 2),
"lower": round(max(0.0, avg - 1.28 * sigma), 2),
"upper": round(avg + 1.28 * sigma, 2),
"lower_95": round(max(0.0, avg - 1.96 * sigma), 2),
"upper_95": round(avg + 1.96 * sigma, 2),
}
for d in future_dates
]
algo = "average"
else:
fcast, lower, upper = _fit_holt(values, FORECAST_MONTHS)
fcast, lower_80, upper_80, lower_95, upper_95, algo = _fit_sarima(values, FORECAST_MONTHS)
forecast_pts = [
{"date": d, "amount": round(max(0, f), 2), "lower": round(l, 2), "upper": round(u, 2)}
for d, f, l, u in zip(future_dates, fcast, lower, upper)
{
"date": d,
"amount": round(max(0.0, f), 2),
"lower": round(l80, 2),
"upper": round(u80, 2),
"lower_95": round(l95, 2),
"upper_95": round(u95, 2),
}
for d, f, l80, u80, l95, u95 in zip(future_dates, fcast, lower_80, upper_80, lower_95, upper_95)
]
results.append({
"category_id": cat_id,
"category_id": str(cat_id),
"category_name": cat_name,
"monthly_avg": round(float(np.mean(values)), 2),
"actuals": actuals[-6:], # last 6 months for display
"algorithm": algo,
"actuals": actuals[-12:], # last 12 months for display
"forecast": forecast_pts,
})
# Sort by monthly_avg descending (highest spend first)
results.sort(key=lambda x: x["monthly_avg"], reverse=True)
return results